Tanith: AI native disaster response case study by Abhishek Paatni Source: https://tanith-intelligence.vercel.app/ This is the full text of the case study page, generated at build time. Tanith Overview # Tanith: disaster response in 90 seconds, not 48 minutes. The whole case study in 30 seconds. An AI generated walkthrough of one 90 second response: detection, verification, coordinated dispatch, ground handoff. Scan in 10 seconds 48 MIN → 90 SEC Target response 3RD / 100+ Global qualification 47 / 50 Top jury score, tied 2nd of 20 9 DAYS Concept to shipped The problem Spotting a disaster is already fast. Coordinating the response is not. The brief named a 48 minute gap between detection and a coordinated response on the ground, and asked for 90 seconds. At Mati, Greece, that gap cost 104 lives. The reframe Most teams redesigned the dashboard. I redesigned what the operator is for. AI is the agent. The human is the judge. What I built Tanith, an AI native command system for critical airspace. The AI detects, verifies and lines up the whole response. The human approves anything that can't be undone. Two states: a quiet Idle overwatch, and a Focused alert view that shrinks the screen to one decision. Every AI call, operator action and failsafe goes into a legal audit log. How the 90 seconds work - T+0s A sensor trips. Three AI layers clear the noise. The quickest drone launches, not the closest. - T+30s Drone in transit. Four feeds spin up: night vision, thermal, 4K, and a map with the trajectory. - T+60s The AI checks the feeds against past wildfires and live weather. Confidence climbs. - T+90s Wildfire confirmed at 94%. The operator proceeds or aborts, or a 25 second failsafe proceeds and logs it. Ground teams are already moving. Four decisions I owned, and what each one cost - 01 A human judge instead of full autonomy. Cost: Seconds of latency on every irreversible action. - 02 A new company instead of the rebrand the brief asked for. Cost: Jury points. I'd make the call again. - 03 Dead Man's Switch: four escalation layers, not one. Cost: Complexity. - 04 Send the quickest drone, not the closest. Cost: Explainability. What I owned Brand, product strategy, information architecture, UX, the 90 second decision chain, and the live Framer site, which I wrote, designed, built and shipped alone. I led a team of four. The read that mattered most "You people have gone way outside the problem statement and truly tried to work out a solution that can actually be called an AI-native solution with human in the loop. Nobody in this designothon or designothon before this ever worked like that." Prathmesh, Lead Product Designer, FlytBase ## The brand and the product at a glance. Tap any image to expand it. Conference signage, built for the screens where govtech contracts get decided. The monogram on hardware. Trust starts at the airframe. Commanding Autonomy in Critical Airspace. Built for emergency operations, surveillance and critical missions. Three messages, three audiences, one identity. The Overwatch state. Drones at rest, heat halos showing regional risk. Silence is the system's default, and its safety signal. A sensor trips at 3:47 AM. The drone is in the air before the operator's screen finishes flashing. Four feeds, one conclusion. 94% certainty and a 25 second countdown. The Dead Man's Switch at work. Three units, three ETAs, one bulk dispatch. The system assembles the response. The operator approves it. ## Detailed case study begins here July 23, 2018 · Mati, Greece ### The fire reached homes before help did. 104 people died, most of them in the first 90 minutes. Tanith is the command system we designed to close that gap. The brief put the average time from detection to a coordinated response at 48 minutes. We designed for 90 seconds, in 9 days, as a team of four that I led. 9 Days 3RD / 100+ Qualification 5TH / 20 Final Jury Use the menu at the top right to jump to any chapter, or follow Up next at the end of each one. Watch the AI concept video ## We didn't redesign the dashboard. We redesigned what the operator is for. Every team got the same brief: take 48 minutes down to 90 seconds. Most of them went to work on the dashboard. We didn't. Legacy command software treats the human as the pilot. They navigate, read the data and decide what comes next. That made sense in 2015. Today, drones can detect, launch and navigate on their own. At 3:47 AM the human isn't a pilot. They are the bottleneck. So we flipped the role. - AI is the agent. It detects, launches, navigates and gathers the evidence. Watching many feeds at once and finding a path in milliseconds are things people are bad at, especially under pressure. - The human is the judge. They check what the AI found, decide whether the confidence is high enough, and approve anything that can't be undone. They carry the legal accountability, and they can overrule the system. Every screen in Tanith follows from that split. The human brain is a decision engine, not a search engine. Pair operations: a senior commander and a supporting officer, the wingman setup our research found in emergency rooms, air traffic control and military command. One tired operator at 3:47 AM is a failure mode, so the system assumes two. ### The harder reframe: a new company, not a rebrand. The brief asked for a FlytBase rebrand. We delivered a new enterprise. A company filed under drone fleet management software doesn't make the bidder lists for defense, govtech and emergency services contracts. Those lists are gated by category, relationships and security posture (full reasoning ). So we built a company that could walk into those rooms, and named it Tanith, after the ancient goddess of safe passage. The jury docked points for that decision. The FlytBase lead designer thanked us for it. ## What I owned. Who I built it with. I led a team of four. We had never worked together before day one. Saniya and Lakshmi were design students. Vanshika was a UI/UX designer about a year into her career. That team shipped Tanith in nine days. - Brand strategy and positioning. The name, the new enterprise reframe, the archetype, the tone of voice. - Product strategy and IA. The Idle and Alert states, the human as judge model, the Dead Man's Switch, the edge case map. - UX flows and decision logic. The 90 second chain, the confidence thresholds from research (Stack 5 ), and quickest drone dispatch. - The Framer website. Written, designed, built and shipped alone, in parallel with everything above. - AI image generation. Atmosphere, pitch material and the future vision shots. - Team leadership. Strategy, settling disagreements, holding scope, and delivering against a nine day clock. Day 2, 5:07 PM. Presenting the gameplan to the team, with the name shortlist on screen (Tanith, Hedron, ORDIN). The rest of the team built against this plan for seven days. ## What we found. Not what we did. ### The problem statement. A single operator must reliably detect, verify, and coordinate life-saving drone actions across multiple zones under extreme time pressure and uncertainty, while the system earns and maintains trust in every second. Nine days left no room for research theater. We built five evidence stacks, each backed by named sources. All five pointed at the same root cause. ### Stack 1: the competition. Every player solves part of the response loop. Nobody owns the whole chain, from detection to a coordinated ground response with a human judging. The gap is the chain itself. Who we looked at - DroneSense. Drone fleet management, no decision intelligence. - Skydio. Defense grade hardware and autonomy. - Paladin Drones. Drone first responder dispatch for public safety. - Auterion. An open source autopilot operating system. - Palantir Gotham. The benchmark for enterprise data fusion. ### Stack 2: how high stakes interfaces work. We studied 15 interface patterns from air traffic control, military command, nuclear control rooms, surgical dashboards and 911 dispatch. They agree on five rules: silence means safe, confidence shows as a range, irreversible actions wait for a held confirmation, alerts sort by severity instead of source, and the operator's mental load is the constraint that matters. Four of the fifteen patterns - STCA (Short Term Conflict Alert), air traffic control. Graduated alerts instead of on and off. - MIL-STD-2525, NATO command and control. A "halo of ambiguity" around uncertain hostile contacts. - Westinghouse Ovation "Shelve", nuclear control. Alarm suppression with a full audit trail. - RapidSOS, 911 dispatch. Incident data passed to first responders with its confidence attached. - Eleven more, each documented. ### Stack 3: how real drone operators read screens. Experienced operators look at the threat feed first, telemetry second and mission status third. Novices do the reverse. Interfaces built on the expert order cut decision time by about 40% in published trials, so we built on it. Sources A Springer eye tracking study comparing novice and experienced UAV pilots. Advexure's field guide to public safety drone operations. An arXiv field trial paper on firefighting with autonomous drones. ### Stack 4: how these systems fail. The biggest failure isn't technical. It's the moment the operator freezes, asked to judge something they aren't ready to judge. In air traffic control simulator studies, pilots routinely ignore low priority AI flags. In a fire, that is lethal. The Dead Man's Switch exists because of this finding. Sources Ten documented failure scenarios, drawn from CAA UK drone safety reports, MDPI studies on vision AI fire detection, NCBI work on GPS spoofing, anvil.so on human error in drone operations, and anikedia research on drone UI overload. ### Stack 5: what makes people trust AI. Operators trust AI when it shows confidence as a range, not a single number. The thresholds we took from the literature: below 65%, the AI asks and the operator decides. From 65% to 84%, the AI suggests and the operator approves. At 85% and above, the AI acts and the operator keeps the override. AI SUGGESTS. Operator approves or rejects. PROCEED Drag the slider. The system only acts on its own when confidence clears 85%. These bands decide whether Proceed is dimmed, highlighted or counting down, and how long each decision holds. Sources CDC crisis communication research, NASA studies on handing off control, EU civil protection coordination protocols, and NTSB findings on collision avoidance. ### The root cause. The operator at 3:47 AM isn't failing because the screens are wrong. They're failing because the role is wrong. They're asked to be pilot, analyst, dispatcher, witness and judge at once. Nobody does all five well in the first minute of an emergency. So the redesign starts with the role, not the screen. Day 3, 6:38 PM. Lakshmi walking the team through the competitor scan, with DroneSense, Paladin Drones and Auterion on screen. Research was something we argued over together, not a solo doc. ## Three options. One choice. The tradeoff named. Every real design decision gives something up. These are the four that shaped Tanith, and what each one cost. ### Who acts: the AI or the human? Option A Full autonomy The AI detects, decides and acts. Humans review after the fact. Roughly where Anduril Lattice is heading. Option B Human driven, AI as support The operator does everything and the AI surfaces information. Where most agencies are today, and roughly what Palantir Gotham serves. Option C The AI investigates and lines up the response. A human approves anything irreversible. The Tanith path. Everything else keeps running on its own, under human oversight. Why not A Liability can't legally or morally sit with an AI. Insurance, evidence chains and post incident audits all need a named human. A regulator can subpoena the operator. They cannot subpoena the model. Why not B This setup is what causes the 48 minute delay the brief asked us to fix. Human driven coordination doesn't scale to supervising autonomous systems. Why C, and what it costs Latency on irreversible actions. Holds are measured in seconds: about 5 for a small call, 25 for a fire dispatch, longer for bigger ones. That buys a human who answers for every irreversible step. We revisit when the law lets an AI be the accountable party. That is not 2026. ### A new company, not a rebrand. The brief asked for a FlytBase rebrand. We weighed three ways to answer it. Option A FlytBase rebrand Same company, new identity, Tanith as a module. Cheapest, and closest to the brief. Option B FlytBase sub brand A product brand inside FlytBase, like "FlytBase Tanith". The middle path. Option C An entirely new enterprise New company, new positioning, new go to market. FlytBase's drone automation stays underneath as infrastructure. Why not A Drone fleet software doesn't appear on the bidder lists for defense, govtech or emergency services contracts. Category is a procurement gate, and a new storefront doesn't change the category in a contracting officer's eyes. Why not B A sub brand inherits the parent's category. To a defense buyer, "FlytBase Tanith" still reads as drone software. Why C, and what it costs Capital. A new enterprise costs far more than a rebrand: brand, positioning, go to market, regulatory posture and security clearances, all from zero. The jury docked us for exactly this, calling it harder to deploy. We held the position. The defense, in one paragraph. The cheap option fails earlier. A company in the wrong category loses every bid it enters. One in the right category competes from day one. The jury's worry about deployability doesn't survive that comparison. The FlytBase lead designer, who works inside that structure every day, saw it. We revisit when drone software vendors can credibly sell AI command intelligence with human oversight directly. Commanding Autonomy in Critical Airspace. The booth a defense procurement officer walks past at a govtech conference. Drone management software doesn't get this booth. ### The Dead Man's Switch: when the operator doesn't respond. Operators go missing for reasons emergencies don't wait for: asleep at 3 AM, in the bathroom, on another call, in shock (Stack 4 ). The system needs a fallback that is neither full autonomy (rejected above ) nor a deadlock. Option A Earthquake chair The chair shakes and sounds an alarm when a critical alert goes unanswered. Option B Smartwatch escalation The alert moves to the operator's wrist when they're away from the desk, with haptics, a flash and sound coded by severity. Option C Peer escalation A free commander in another zone gets control and joins the comms loop. Option D Calibrated auto proceed No response within a window sized to the decision? The AI takes the recommended action and logs the failsafe. Fatal actions only auto proceed above 85% confidence (Stack 5 ). Why not just one Each one fixes one failure and leaves the rest open. The chair fails if the operator is in shock at the desk. The watch fails with no signal. Peer escalation fails when nobody is free. What we chose All four, stacked. D is the floor that always happens. A, B and C sit above it: the chair, then the watch, then a peer, and only if all of them fail, the auto proceed. What it costs Complexity. Four layers are more to design, test and break. Picking one would mean accepting a known failure as the default. In this product, we wouldn't. Layer 2 of the Dead Man's Switch: the wrist alert when the operator is away from the desk. ### Send the quickest drone, not the closest. The usual answer is the closest drone. Tanith sends the one that will arrive first. They are often different. A drone already in the air can turn and get there faster than a docked one that is nearer but has to undock, climb and speed up. Against a 90 second target, that gap can be 15 to 40 seconds. In a wildfire, it's the difference between catching the fire 50 m from homes and 200 m from them. The math is the AI's job. The operator sees which drone is going, its ETA, and why, if they ask. The eight variables, checked for every drone in under a second - Altitude and heading. - Vertical and horizontal speed. - Distance from dock, for docked drones. - Terrain between each drone and the site. - Weather between each drone and the site. - Fuel remaining. - Sensor payload match for the type of anomaly. No human can weigh eight variables across twelve drones in under a second. The AI can. That is what AI native dispatch means. What it costs: explainability When the AI skips the obvious drone, the operator may not see why. So the screen shows both ETAs: the closest drone's and the chosen one's. What the AI weighs for each drone: altitude, vertical and horizontal speed, latency, distance from dock. Hover any drone to see them. ## What we built. The system first, then the screens. Here is how Tanith is organized, then one incident followed through it. ### Two states, and a funnel between them. In Idle, the map is calm on purpose: drones docked or on patrol, risk halos over regions, live weather. Calm is the safety signal, so when something changes, it means something. In Alert, the map collapses into the Focused screen: four feeds, the AI's assessment, and one decision. Between them sits the Funnel of Truth. Three AI layers filter false positives, match the event against past anomalies, and check confidence against the trust thresholds. Most of what reaches a sensor never reaches the human. The human is too expensive to spend on noise. Underneath it all runs a legal audit layer that timestamps every telemetry frame, AI recommendation, operator action and auto proceed. A system this autonomous can't ship without one. ### The information architecture. Idle, then Alert in two phases: assess, then respond. The screens inside each, and two layers that never switch off. Tanith command console - Idle state #### Overwatch - Overwatch map: Drones docked or on patrol, risk halos by region, live weather. - Drone detail, on hover: Altitude, speed, distance from dock, battery, live feed. - Sector status: Monitored, unmonitored, or covered by a patrol drone. - Alert state · Assess #### Focused screen - Four live feeds: Night vision, thermal, 4K camera, map with trajectory. - AI verdict: Anomaly type, with confidence shown as a range. - Situational report, on demand: Wind, temperature, humidity, visibility, pressure. - Decision bar: Proceed, Abort, and the hold timer. - Alert state · Respond #### Dispatch and coordination - Deployment panel: AI strategy, suggested units with ETAs, Dispatch All. - Spread projection: Predicted path, zones and homes at risk. - Ground team channel: Routes, ETAs, comms per team or all teams. - Active response: Drone feeds to crews, evacuation guidance by light and speaker. - Idle → Assess: An event clears all three AI filters. - Assess → Respond: Proceed, by the operator or the failsafe. - Assess → Idle: Abort. The drone flies back to its dock. - Respond → Idle: The incident closes and is archived. Always on Legal audit layer. Every AI call, operator action and failsafe, timestamped and archived. Dead Man's Switch. Chair, then watch, then a peer, then a calibrated auto proceed (why ). The full flow map from the sprint. Passive state, focused screen, situational report, deployment, path projection, ground coordination, active response. Abort sends the drone back to dock. Legal audits run under everything. ### The 90 second chain. Four beats. Each one is a screen. T+0s Sensor trip. What happens Edge sensor 42 spots a thermal anomaly in Sector 04. The event clears the Funnel of Truth, and the AI launches the quickest drone in the region, not the closest. Idle flips to Alert and the sector turns red. DRONE 42A is already chosen. The operator never has to pick a drone. T+30s Drone in transit. Four feeds spin up. What happens Drone 42A is airborne. The Focused screen opens with four feeds: night vision, thermal, 4K camera, and a map with the trajectory. Nobody asks the operator anything yet. The system is bringing them the evidence. Feeds live as Drone 42A reaches the north west perimeter at 34.02, -118.45. The map quadrant shows trajectory and ETA. T+60s AI assesses. Confidence climbs. Thermal and video confirm the heat. The AI checks it against past wildfires and live weather (wind 13.4 m/s). Confidence passes 65%, then 75%, then 85%, where Proceed unlocks. T+90s Wildfire confirmed at 94%. Proceed lights up and a 25 second hold starts. The operator can proceed early, abort, or let the timer run out. ### The held moment. Every irreversible action in Tanith passes through this screen. What each part of the screen does - "Wildfire Detected" is the AI's conclusion, not its inputs. - "94% Certainty" shows as a number and as a range bar, because operators trust ranges more than absolutes. - "PROCEED" is highlighted because confidence is above 85%. - "The Mission will automatically proceed in 25 sec" is the hold, sized to this decision. - "ABORT MISSION" is always available and never times out. - The feeds stay visible behind it, so the operator can check the evidence in the seconds they have. The hold scales with the stakes. 25 seconds lets an experienced operator glance at four feeds and confirm. Confirming a noise filter might hold for 5 seconds. Ordering an evacuation might hold for 60 or more. Need more context? The AI situational report opens next to the countdown without pausing it. The high risk variant. Wind 45.2 km/h, air 48°C, humidity 12%, visibility 1.5 km, pressure 1013 hPa, tagged CRITICAL. The decision comes first. The evidence is one tap away. If nobody acts in 25 seconds, Tanith proceeds, logs the failsafe and records every input behind it. That is the floor of the Dead Man's Switch . ### Handing off to the ground. Once the operator (or the failsafe) proceeds, Tanith stops assessing and starts responding. What each part of the panel does - "PROCEED TO DISPATCH" keeps the same action verb across the whole chain. - "Incident ID: WF-0428" gives every incident a reference for the audit trail. - "AI Strategy: Tier 2 Fire Suppression + Zone 4 Evacuation" states the plan in plain language before any unit moves. - Three ready units. Fire Suppression, Engine Co. 12, Station 04, ETA 04:12. Civil Defense Unit 09, ETA 06:45, for The Oaks residential block (14 houses). EMS Unit 05, ETA 05:20, staging for heat distress. - "Dispatch All Teams" is one bulk action, because operators do best with two or three strong choices, not long lists. Three units, three roles, three ETAs, one Dispatch All. The AI assembled the response. The operator approves it. Crews get everything on their tablets: GPS, the projected spread, approach routes, wind, evacuation status and comms channels. Nobody arrives blind. The receiving end. The AI saves crews the 8 minutes they would spend radioing for basics. The operator's 90 second call becomes the crew's 8 minute head start. ### Plugging into real sensors and drones. Designed against real hardware and open standards. Not wired to live hardware in nine days. - In. Tower thermal cameras that detect on the device, satellite hotspot feeds (NOAA, NASA FIRMS), ground smoke and weather sensors. Each becomes one standard event, so a new sensor needs an adapter, not a redesign. - Fly. Dock based drones through FlytBase's existing fleet layer. Tanith picks the drone and the mission, and FlytBase flies it. Other fleets connect over MAVLink. - Out. Dispatch goes into the agency's existing dispatch software as standard CAP alerts, and to crew tablets. Tanith feeds the dispatch center. It doesn't replace it. An agency starts with the cameras and satellite feeds it already has, then adds docks where risk runs highest. Partial coverage still works. It just verifies slower. ### When things fail. A system this autonomous is judged by how it fails. The rule: degrade toward a human, never toward silence. Every failure is shown to the operator, has a fallback, and is logged. When a sensor fails - Sensor goes dark. The sector is marked unmonitored and the nearest patrol drone covers the gap. It shows as a status change, not an alarm. - False positive. The funnel filters most. If a drone disproves one, the alert closes with a logged reason. - Sources disagree. Nothing auto proceeds until a drone has looked. The drone goes anyway, because a look is cheap. When a drone fails - Link lost. The drone returns on its last safe path, and the next quickest drone launches at once. - Low battery or high wind. A drone that can't finish safely is never picked. If none can fly, cameras and satellite verify, and the operator is told why. - GPS spoofed. Position is cross checked on board. On a mismatch the drone returns and a security incident is logged. When the AI is unsure - Below the 85% threshold. Proceed isn't highlighted and fatal actions never auto proceed (Dead Man's Switch ). The system asks for a second look and the operator keeps deciding. When the human fails - No response. The Dead Man's Switch takes over. - Wrong call. An abort is never final: the drone stays on station and alerts a peer if confidence climbs. A wrong proceed can be recalled in one action. When the system fails - Console or network down. A standby console takes over, drones in the air finish on board, and alerts reroute to peers and smartwatches. It degrades to manual. It never goes quiet. That is what 90 seconds buys: a coordinated response already moving while legacy systems are still gathering their first sitrep. ## User journey: one fire, one night shift. The duty commander, 3:47 AM. A wildfire starts in Sector 04. What they do, what Tanith does, and how it feels, next to how the same night goes today. - Before T+0 ### Overwatch Operator does: Watches a quiet map. Nothing moves unless it matters. Tanith does: Filters sensor noise through three AI layers and stays silent. Without Tanith: Alarms all night. Real alerts get lost in the noise. Operator feels: Calm - T+0s ### Alert Operator does: Sees the screen flip to Alert. One sector turns red. Tanith does: The event cleared all three filters. The quickest drone is already flying. Without Tanith: 6 min before a human even hears about it. Operator feels: Startled - T+30s to T+60s ### Verify Operator does: Watches four live feeds: night vision, thermal, 4K, map. Tanith does: Checks the feeds against past wildfires and live weather. Confidence climbs to 85%. Without Tanith: 14 min of calls and cross checks. Operator feels: Unsure, then steadier - T+90s ### Decide Operator does: Proceed or Abort. One decision on screen. Tanith does: 94% confidence. Proceed highlighted. A 25 second hold, then the failsafe. Without Tanith: 5 min to decide, often alone, often frozen. Operator feels: Peak pressure - T+90s onward ### Dispatch Operator does: One action: Dispatch All. Tanith does: Built the plan while the drone flew. 3 units, ETAs 04:12, 05:20, 06:45. Without Tanith: 18 min coordinating units by radio, then 5 min to dispatch. Operator feels: Back in control - After dispatch ### Handoff Operator does: Talks to crews, monitors, closes the incident. Tanith does: Loads crew tablets with GPS, spread map and routes. Logs every action for audit. Without Tanith: Crews arrive blind and spend 8 min radioing for basics. Operator feels: Calm again "Without Tanith" splits the brief's 48 minutes by our model (see the projections ). Tanith's times are design targets, not measurements. The stress peak doesn't disappear. It shrinks, from 48 minutes of chasing information to one 25 second decision with the evidence on screen. ## Projected, not measured. Tanith was never deployed, so nothing here is measured. These are projections and test targets, labeled as such. ~3 Projected minutes, expected case −93% Projected vs the 48 min baseline 8 Projected minutes, worst case ### Projected response time. The brief's 48 minutes, split into stages, then run through Tanith. ``` STAGE LEGACY TANITH (brief) (projected) Detection reaches a human 06:00 00:10 Verification 14:00 02:30 Coordination 18:00 in parallel Decision 05:00 00:25 Dispatch 05:00 00:10 TOTAL 48:00 03:15 ``` The saving doesn't come from one faster step. Legacy response verifies, then coordinates, then decides. Tanith does all three at once: the drone launches at the alert, and the dispatch plan is built while it flies. The assumptions - The 48 minutes is the brief's figure. The stage split is our model. - Both columns stop at dispatch. Travel time for ground teams isn't counted. - The nearest docked drone is about 2 km out, cruising at about 60 km/h. - The operator uses the full 25 second hold. ### Operators under pressure. Not yet tested with real operators (what's next ). What exists is a design aimed at how pressure breaks people, and a test to prove it. - Tunnel vision. One decision on screen at a time. - Freezing (Stack 4 ). A held deadline, then the Dead Man's Switch. - Alarm fatigue. Noise is filtered out. Idle stays silent. - Choice overload. Three ranked units, one Dispatch All. - Distrust of the AI. Confidence as a range, and the closest drone's ETA beside the chosen one. ### The pressure test. Two rounds of six dispatch operators on a simulated night shift, with four scenarios: a real fire, a false alarm, a lost drone, and a second alert in the middle of a decision. ``` MEASURE PASS MARK Time to decision after alert ≤ 25 s Correct calls (proceed / abort) ≥ 95% Missed critical alerts 0 SUS usability score ≥ 80 (industry avg 68) Follows the AI on low confidence < 10% ``` The last row matters most. Operators should trust the AI when it is sure and question it when it isn't. ### From projection to measurement. A 30 day shadow pilot with one agency. Tanith runs silently beside the real dispatch center, and the agency's own response times are the control. ## A brand that opens doors before the product does. Because Tanith became a new company (why ), the brand had to do work the product couldn't yet. In three seconds, a defense procurement officer had to think "serious vendor", not "student competition project". Six choices got us there. None of them were about taste. Visual direction Typography Color The monogram ### The name. Tanith, the Phoenician goddess of safe passage, chief deity of Carthage, invoked when caravans crossed dangerous ground. Invented names read like agency output. A name with two thousand years of meaning reads as inherited authority. It ends on a hard consonant that lands when spoken, and it doesn't clash with any aerospace or defense trademark in the markets that matter. The two finalists that lost - Hedron. Invented and easy to read, but it suggested geometry more than gravity. A software name, not a defense one. - ORDIN. Too close to "ordinance" without a clear meaning of its own. It risked sounding like a procurement category instead of a company. Tanith won because it already meant something. The others had to invent it. ### The monogram. Three stacked tiers: a chevron on top, two arrows in the middle, a diamond at the base. Read top to bottom, it is the AI watching from the sky, the operator's decisions in transit, and the ground the system protects. It is also a drone's view, looking straight down. The full exploration. Ten variants, reference scans of Spectra, Airspace and other defense identities, sketches and refinements. The stacked chevron and diamond locked on iteration 8. Tanith Full lockup. Custom letterforms paired with the monogram. ### Color. Indigo, #4927B0. Blue is what every defense and enterprise brand already uses. Red is taken inside the product, where it means danger. Indigo is neither. Why not blue, why not red - Blue is the category default. Walk a govtech conference and twenty booths in a row are some shade of it. Standing apart cost nothing, so we took it. - Red already marks critical alerts, aborts and hazard zones in the UI. A red brand on a red alert screen would make an operator wonder whether something is the logo or on fire. - Indigo is cool and authoritative without being blue. It nods to Tyrian purple, the imperial color of the Phoenician world, it holds up on the dark UI, and it keeps the cool trust register defense buyers expect. ### Typography. Three typefaces, one job each. - Space Mono for display, 48px and up. Monospace reads like instruments: air traffic control, telex, terminal output. The cockpit, not the marketing deck. - Inter for UI, under 48px. Legible at small sizes on screens, for an operator reading at 3 AM in a dark room, sometimes through tears. - Source Code Pro for data and telemetry. Zero never looks like O, and one never looks like l. That matters when you read sensor IDs, GPS coordinates and timestamps. ### Visual direction. The brand had to say three things: speed, autonomy, authority. Everything else got cut. The rules we locked - Graphics. Simple geometry only: map lines, circles, arcs, cones, straight lines. Instruments, not ornament. - Illustration. Clean vector diagrams, top down or isometric, with no characters. Only for explainers, decks and onboarding. - Motion. Straight or curved paths only. No bounce, no elastic. Things slow down like heavy objects, not friendly mascots. - Materials. Matte metal, the language of aircraft and critical infrastructure. Subtle grain, so it never looks sterile. Directional depth to guide attention without panic. One soft glow, kept only for the AI's active state. ### Touchpoints. Software is one surface. Hardware, wearables, print and signage are the rest. Drone airframe. Monogram and accent stripes. A brand built for hardware. Operator station boot screen. "Command. Decide. Act." Conference LED. "Autonomous Command Systems. When Every Second Matters." Govtech expo booth. "Commanding Autonomy in Critical Airspace." Three poster gallery wall. Three messages, three audiences, one identity. Public billboard at night. The monogram and the category, "Disaster Response System". Operator wearable. Layer 2 of the Dead Man's Switch. Personnel ID. The brand on a body, not just a screen. Operator uniform. The brand goes where the crew goes. Field cap. Same logic. The brand follows the person. Headquarters signage. The brand at home. Building exterior. The brand at city scale. Day 3, 7:21 PM. Lakshmi building the final monogram in Illustrator, vector by vector, against the visual rules above. ### Brand archetype Hero, with a Sage influence. A hero's posture, the protector who acts when others can't. A sage's voice: calm, evidence led, never shouting. Defense buyers and emergency service directors take that counterpart seriously. ## The website. Written, designed, built and shipped solo. ### tanith.framer.website When Every Second Counts, Tanith Responds. Visit the live site The live site at tanith.framer.website was the day nine deliverable. I built it alone in Framer: copy, design, layout and deployment. Why solo: a brand voice needs one editor. Built by committee in nine days, the voice and rhythm fracture. Tanith sounds calm, evidence led, and never longer than it needs to be. That took one person holding the pen. Why Framer: design fidelity and deployment speed in one tool, against a hard submission deadline. Its job: read like a vendor a defense contracting officer could route an RFP through. A company, not a portfolio piece. The blocks on the page, and what each one is for - Hero. "When Every Second Counts, Tanith Responds." The category claim and the difference, in one sentence. - About Us. "Born From Disaster. Built To Protect." Why this company exists. - Our Aim. Purpose, mission and vision, the institutional voice govtech buyers expect from a vendor. HR readers can skip it. Procurement readers can't. - How It Works. Detect, Confirm, Act. The 90 second chain in three beats for someone scanning. - Scenarios. "Operator Centered, Not AI Obsessed." Four product UI tiles, each with a one line principle: feed and context for fast decisions, a confidence based call to action, the overwatch state, live comms. - Why Us. Six principles, one sentence each: Faster Than Human Reflexes, Clarity Over Complexity, Confidence Transparency, Built for Real Environments, Human In The Loop Control, From Detection to Delivery. - Testimonials. Four positioning quotes from four personas in four regions: California fire ops, a Berlin hospital ER, an emergency drone unit operator, Japan flood response. Each one speaks to a buyer segment. - Solution video. The same AI generated walkthrough that opens this case study. - Onboarding. The 30 day rollout: site audit, operator training, live deployment. Procurement needs to know what buying this looks like. - FAQ. The six questions a procurement officer or head of operations asks before signing anything. Every block answers a buyer's question. Process pages, designer credits, methodology and color palettes stayed off the site. They live here instead. The case study is for hiring managers. The website is for buyers. ## What the jury said. What it cost. What I'd do again. ### The numbers. ``` Qualification stage: 3rd of 100+ teams globally. Final jury (top 20): 5th of 20. Final score: 82 / 100 (Rajat 35 / Nisha 47) Winning team (KRANTI): 93 / 100 (Rajat 44 / Nisha 49) ``` The final scoring sheet from the jury: top 20 teams, two judges, out of 100. Certificate of Merit, Design Fusion Season 3, signed by Vannsh Agrawal, founder and CEO of Designare, for the top 20 finish. ### On the record. "The solution, UI, flows, IA and thinking behind the brand and product looks grounded in reality. I feel like this is something that can be shipped to developers for development and then shipped to the world, and I loved the fact that you people did not try to make it pixel perfect and gimmicky." Rajat Patel, Jury Member, Design Fusion Season 3 "I never expected any team from the designothon to ever come up with a solution that can be pitched in real life to investors by navigating real world regulations, laws and constraints while using the current UI, flows and design system of FlytBase. You people have gone way outside the problem statement and truly tried to work out a solution that can actually be called an AI-native solution with human in the loop. Nobody in this designothon or designothon before this ever worked like that." Prathmesh, Lead Product Designer, FlytBase(quote used with permission) Nisha Kumari's read is about the pitch, so it sits with the pitch below. ### The honest read on 5th place. Nisha scored us 47 of 50, tied for her second highest score in the top 20. Only the winners, KRANTI, scored higher with her (49). On storytelling and clarity, we were effectively in the top two. Rajat scored us 35 of 50, below average for the top group. He told us his lens included whether the solution could realistically deploy, and the new company decision (why we made it ) added cost to deploying it. Would I make the same call again? Yes. We knew the decision would cost us and held it anyway. Drone management software doesn't make the procurement lists this product needs, and a company in the right category is cheaper than one that loses every bid. So 5th place is the price of a position the jury didn't fully accept and the FlytBase lead designer did. Operators, engineers and product leaders judge substance. Competition juries judge fit to the brief. I think that trade favors the job I'm after. ## Two minutes. Three disasters. One solution. The final jury round of Design Fusion Season 3 was a live pitch of exactly two minutes. Run over and you lose points. Run under and you lose conviction. Nine days of work had to fit in less time than most TikToks. ### How it opened: three real disasters, one beat each. Mati, Greece, 2018. 104 dead, most in the first 90 minutes. Officials were later convicted for failures in the response. Pedrógão Grande, Portugal, 2017. 66 dead on a single road, trapped as the fire crossed their evacuation route. It became known as "The Road of Death." Black Summer, Australia, 2019 to 2020. 33 dead, 3 billion animals killed, 24 million hectares burned. The largest fire event in Australia's recorded history. Three countries, one pattern: the response was slower than the fire. ### How it built. Then the chain: detection, AI assessment, the operator's call, coordinated dispatch, crews moving. The same 90 seconds from the solution , told as a wildfire the jury could feel happening. Then the brand, the deck and the close. Two minutes flat, and no questions needed. ### What the jury said. "The storytelling of the pitch to explain a scenario that not only tells us the impact of the work you did but also shows how you were actually able to solve the 45-minute rescue delay to 90-second response, alert and rescue operation, keeping the pitch under 2 minutes but exact 2 minutes was amazing. Nobody has been able to do that and the fact that I do not have to ask any question to understand your solution is also wonderful." Nisha Kumari, Jury Member, Design Fusion Season 3 This quote means more to me than the placement. Squeezing a whole system into 120 seconds without losing the story or the rigor is hard. Nisha is saying we kept both, clearly enough that nobody on the panel needed to ask a question. ## The dent Tanith was designed to make. Third party data only. Every figure is sourced, and where there is an estimate, the method is named. ### Detection is fast. Response isn't. NOAA's Next Generation Fire System, live since February 2025 and used by 90% of National Weather Service forecast offices, can spot a new wildfire from satellite imagery and alert authorities in as little as one minute [Source: NOAA, May 2025]. In Oklahoma's 2025 Great Plains outbreak, up to 80% of satellite detected fires reached authorities before any 911 call. Coordinated response is another story. A peer reviewed study of the 2020 California fire season matched 7,214 wildfires against real aircraft flight data (ADS-B): ``` Average aerial response time (CAL FIRE): 700 minutes (≈ 11.7 hours) Median aerial response time: 751 minutes (≈ 12.5 hours) Fires with response within 20 min: 55 of 7,214 (0.76%) Fires with NO aerial response in 24 hours: 5,737 of 7,214 (79.5%) ``` [Source: Magstadt et al., International Journal of Disaster Risk Reduction, Vol. 137, 2026.] That leaves a gap of about 699 minutes per incident between detection and aerial response. Nearly 80% of California wildfires got no aerial response at all in the first 24 hours. Tanith's target, and the brief's, is 90 seconds. That is the dent. ### Why minutes matter. A 2024 analysis of California wildfire costs, published by the Moore Foundation, found: - Responding one hour sooner cuts the number of large fires by 16%. - Responding one hour later makes it 29% more likely that a fire burns more than 20 hectares. - Cutting average response by 15 minutes would be worth "billions" in economic benefit. [Sources: Dryad analysis citing Moore Foundation report; California Council on Science & Technology, 2025.] Tanith aims to turn a 700 minute response into 90 seconds. Even if real conditions stretched that to five minutes, it would still be about 140 times faster than today's average. Mati, seen through this data The Mati fire of July 23, 2018 killed 104 people and injured 172. Most died in their cars or homes in the first hours. In 2024, Greek courts found fire service and civil protection officials criminally negligent: drivers were sent toward the flames, evacuation orders came late, and the response was poorly coordinated. Survivor groups put the ruling plainly, as the 2025 anniversary coverage confirmed: "Most of the lives lost could have been saved." Not with better equipment. With a faster, better coordinated response. I'm not claiming Tanith would have saved 104 lives. It didn't exist, and that hypothetical can't be computed. But its architecture (AI detection, AI coordination, a human judge, a ground handoff in under 90 seconds) targets exactly the failure Greek courts spent six years documenting. The scale of the problem California alone had 8,213 wildfires with valid discovery data in 2020. The 2025 Los Angeles fires caused an estimated $164 to $275 billion in economic losses [Sources: PreventionWeb, UNDRR, arXiv]. California spends $1.13 billion a year on wildfire emergency response. Greece lost 104 people in twelve hours in one town in 2018. Australia lost 33 people and 3 billion animals in the 2019 to 2020 season. Portugal lost 66 people on one evacuation road in 2017. Climate models expect fires to grow more frequent and more intense through the 2030s and 2040s. If one regional agency, a CAL FIRE district or a Greek civil protection prefecture, adopted this architecture and closed even part of the gap, the impact would be counted in lives and dollars, not design awards. ### What the dent looks like. The AI drone arrives first with intelligence. The manned helicopter follows, already briefed. The city scale endpoint. A protective layer nobody notices until it is needed. Post incident inspection. The legal audit layer becomes the tool for reviewing what happened. The system in motion. Dock to anomaly zone by the quickest path available. ## Still being built. Tanith is a competition stage prototype with a full brand system, a live brand site, a deep research base and a clear architecture. The next steps are named: a pressure test with real dispatch operators, then a 30 day shadow pilot (the plan ). ### The one thing I'd change. We did zero primary research with real disaster operators. The findings on operator behavior (Stack 3 ) came from published studies. They are good findings, but they aren't a week inside a CAL FIRE dispatch center. If Tanith were a real venture, that would be the next sprint. ### Credits and contact. PROJECT Tanith, AI native command intelligence for autonomous emergency response COMPETITION FlytBase × Design Fusion Season 3, 2025 to 2026 DURATION 9 days, December 2025 to January 2026. TEAM 4 (Chaos Crew) Abhishek Paatni Lead. Brand, product strategy, IA, UX, Framer site. Vanshika Kandya UI screen execution. Saniya Jadhav Research support, secondary scans. Lakshmikanthraju HP Motion graphics, logo refinement. LIVE SITE tanith.framer.website CASE STUDY https://tanith-intelligence.vercel.app/ CONCEPT VIDEO youtu.be/9ZaovEzCepA EMAIL thepaatni@gmail.com PORTFOLIO paatni.vercel.app SCHEDULE topmate.io/thepaatni/1772841 (block a discovery call) Tools Figma, Framer, Adobe Illustrator, Notion, Perplexity, Midjourney, Canva, Claude, ## Nine days. Four people. Whatever sleep was left over. The work as it actually looked: Google Meet windows at 11 PM, the silence after a hard call, waiting for someone's frozen video to come back. Day 1, 11:40 PM. Hands on faces, hands on chins. The first working call where the scope of what we'd agreed to started to feel real. Day 2, 12:49 AM. Lakshmi on a Krishna avatar because his camera had decided enough was enough. We kept going. Day 2, 7:44 PM. Presenting the gameplan, with the name list on screen: Tanith, Hedron, ORDIN. Three candidates. One survived. Vanshika, full focus. The UI screens in the Solution chapter went from sketch to shipped in her hands across four nights. Day 2, 5:07 PM. Strategy review. Three of four on the call, in front of bookshelves, a balcony, a blank wall. The work happens wherever it happens. Day 2, 8:27 PM. Heads down. Different rooms, different cities, one document open in everyone's tab. Day 4, 8:24 PM. Three of four, mid decision. The middle of a project is where teams either pull together or split. We pulled together. If you have done a competition like this, you know it doesn't look like the screens in the solution . It looks like these screenshots. Four people in four rooms, two hours into a call, arguing over whether the confidence threshold should be 85% or 80% while someone tries to remember if they ate lunch. Vanshika ran the screen execution under brutal time pressure. Saniya carried part of the research and helped work out the trust mechanics. Lakshmi held the motion and the final logo passes when there were no hours left. I led it, and I'm proud of how. I'm prouder of who was in the room. Up next Reframe & Role Back to portfolio