Affinda AI Innovation Challenge · Track 3
Nobody confirmed a first aider was sent.
2pm Saturday, 38°C, radio traffic doubled. Somewhere in that traffic someone asked for a first aider at the Lawn Stage, and nobody ever answered. Mo, the safety lead, cannot hold that in their head while handling everything else. Ground Control catches it.
Nothing gets forgotten
Every report that needs a response stays open until someone confirms on the radio that it has been dealt with. Today that list lives in Mo's head.
One decision at a time
Not a chatbot and not a dashboard. Mo's phone shows one card: the most urgent thing that needs Mo. Twelve words, one tap, and the exact words that would be sent.
Ears before eyes
Mo is on foot with an earpiece. Life-safety cards speak once, and "Brief me" gives three items in about ten seconds.
How every report is grouped
One label used to do three jobs. Now the AI fills in three fields for every report: what happened, who should respond, and how fast. The type sets a minimum urgency the AI can raise but never lower. Only a person lowers it.
| Type | Team | At least | Raised when |
|---|---|---|---|
| Medical | First aid | Amber | red if collapsed, unconscious, not breathing or bleeding heavily |
| Missing or found child | Child-safe team | Red | always red |
| Crowd pressure | Security | Amber | red if people can't move, are pushed against barriers or fences |
| Aggression | Security | Amber | red if a weapon or injury is involved |
| Hazard or weather | Site crew | Amber | red if structures, power or lightning put people at risk |
| Staffing gap | Volunteer coordinator | Green | amber at a safety post (first aid, gates, Kids Zone) |
| Volunteer welfare | Volunteer coordinator | Amber | red if the volunteer is unwell |
| Facilities | Site crew | Green | amber if it blocks an exit or water |
| Lost property | Info Tent | Green | never above green unless medication |
Mo can always change his mind
- Reclassify: change the type, team or urgency.
- Reassign: send someone different. The previous responder is stood down in the same call.
- Reopen: bring back something marked resolved.
- Split or join: undo or make a same-incident decision.
- Correct: a radio call can't be unsent, so the AI drafts a correction for Mo to approve.
From any card's "Change or skip", by tapping an open call, or by tapping any line in the log. Every change is logged next to what the AI originally said.
A live map on Mo's phone
The phone shows the whole site at the same size on every screen, with the one decision underneath: every open call as an icon in its urgency colour, and each area's headcount against the roster (First Aid 2/4 when two first aiders don't turn up), because 300 names would be unreadable. Only someone moving between areas is drawn by name, with metres and minutes to go, and the counts change as they leave and arrive. On a decision, the person it would send is shown too. The larger map stays on the control-tent screen.
Movement is simulated GPS in this prototype. On the day it would come from volunteers' phones, shared only while they are on shift.
Where the AI works, and where it stops
The AI does the reading
- Reads each transmission: what, where, how serious, who is responding
- Decides new, update, duplicate or nothing
- Tracks each incident's status from what is said on air
- Links reports in different words (Leo / Leon, red cap / red hat)
- Notices patterns: three heat cases at one stage
- Drafts the next action, naming real people with the right certificate
- Writes the ten-second spoken brief
Ordinary code does the counting
- Waiting-too-long timers per severity (a timer, not a judgement)
- Pattern backstop: a count in a time window
- Card ranking and when to interrupt
- Validation of every model output against the open incidents
- Policy guards: low-confidence and child joins go to a person
- Retiring drafts that are overtaken on air
A person decides
- Sending anyone anywhere
- Releasing a child to an adult
- Stopping a stage or holding entry
- Calling 000
- Messaging any volunteer
- Confirming two reports are the same incident
Every suggestion and every decision is logged with a time and who decided. If a model call fails or returns bad output, the transmission stays in the ledger as unprocessed and becomes a card. Nothing is dropped.
The scripted Saturday
Twenty-eight transmissions over sixteen minutes, written to test the hard parts. A small check script plays them through the pipeline and scores how many land in the right incident, so we can see whether a prompt change helped or hurt. Ground truth never goes to the model.
- T03The Lawn Stage call nobody answers. A radio check arrives fifteen seconds later and buries it.
- T12A second report of the River Stage collapse, from Bar 2, in different words.
- T15"Spilling towards Gate B" is an update at Water Station 2, not a new incident at Gate B.
- T16A man fainting by the bar at the Lawn: the false-merge trap next to the girl at the barrier.
- T17"Leon or Leo, red hat": a found child who is probably the missing Leo. A person confirms.
- T21"Okay to hand him over?" A bright line, drafted as a question with an ID check.
- T22The third heat case at the River Stage in under seven minutes becomes one pattern.
- T24, T26Closing one incident at a site must not close another at the same site.
Some transmissions depend on Mo. Approve sending Sam to the Lawn Stage and Sam reports on scene a minute later; ignore it, and nine minutes in the stage calls again, still waiting.
Open decisions, and the placeholders we chose
These belong to Fieldday, not to software. Each lives in one file (lib/policy.ts) and the whole app follows it.
- How long can each severity wait without a responder?
- Red 60s to acknowledge and 2 min to name a responder; amber 3 and 5 min; green 15 min.
- What makes an incident urgent, and what order do cards reach Mo?
- Each type has a minimum urgency (red, amber or green) that the AI can raise but not lower. Cards are ordered by urgency first, then anything left waiting too long, then the kind of decision (same-incident checks and patterns before ordinary drafts), then age. One card at a time.
- Which actions may the AI draft, and which should it never suggest?
- Drafts: radio calls, texts to named volunteers, move requests and MC announcements. Never: dispatching, stopping a stage, holding entry, releasing a child, calling 000, messaging anyone.
- How should a card read in two seconds?
- Twelve words or fewer, phrased as a question Mo can answer with one tap, with the exact words that would be sent underneath.
- When does the system interrupt Mo?
- Life safety, patterns, and medical or child calls left waiting break into the earpiece once. Everything else waits on the phone or for "Brief me".
- False alarms or silence?
- Lean toward false alarms for children and life safety and toward silence for everything else. A wrong merge is treated as the worst error, because it hides an incident.
Every question in the brief
What we built, what we designed but did not build, and what we left out on purpose.
Awareness
Capture every report
- Channels
- Partly built
- Radio first, because that is where Mo already is. The pipeline takes any text with a sender, so texts and walk-ups typed in by Control use the same path (try the inject box). Group chat, security and first aid feeds are the same shape and come next.
- Radio transcription
- Designed
- Next step: play the script as noisy radio audio and transcribe it. Uncertain words stay marked as uncertain in the transcript and the model is told to treat them as unsure, which lowers confidence and pushes joins to "possibly the same".
- Volunteer reports
- Designed
- One tap plus a voice note, sent as a text. No app install. A first-timer should never fill in a form during an incident.
- Minimum facts
- Partly built
- Every incident records what, where, how bad and who is on it. An incident with no responder is chased by the stall timer, with a drafted call that asks for the missing fact.
- Location
- Built
- A site list with aliases ("the big tree" is Gate B, "dodgems" is the Kids Zone). A direction ("spilling towards Gate B") is not a location. In the script, T15 tests this. People's positions come from simulated GPS: they walk from their post to wherever they are sent, so the map and Mo's phone show who is heading where, how far away and when they will arrive.
- Silence
- Designed
- Each critical post gets an expected check-in interval; a post that goes quiet becomes an incident of its own on the same stall timer.
Merge into one incident
- Duplicates
- Built
- The model compares each report against the open incidents. In the script, Bar 2's "someone's gone down near the River Stage" joins the collapse report (T12) and is not counted as a second heat case.
- False merges
- Built
- Same place and similar words is not enough: the model checks the person, the exact spot and the time. A join below 70% confidence becomes "possibly the same as" for a person to confirm. T16 (a man fainting by the bar at the Lawn) is the trap next to T03 (a girl at the Lawn barrier).
- Linking
- Built
- A found child is never joined to a missing child by the AI. It opens as its own incident linked "possibly the same", and Mo confirms with one tap (T09 and T17: "Leo, red cap" and "Leon or Leo, red hat").
- Updates
- Built
- Status comes only from what is said on air: reported, acknowledged, responder going, on scene, resolved. The model is told never to infer a state nobody said.
- Patterns
- Built
- Two layers. The model can flag a pattern, and ordinary code backs it up: 3 heat or medical cases at one site within 15 minutes. The pattern becomes one card with a drafted stage announcement.
- Jargon
- Built
- Fieldday owns a glossary (Delta, Code Kilo, shirty) that goes into every model call. The team edits it in the scenario file; nobody retrains anything.
Surface what matters first
- Severity
- Built
- Three separate fields for every report: the type (what happened), the team (who should respond) and the urgency (red, amber or green). The type sets a minimum urgency; the AI can raise it from the details but never lower it, and only a person can. A missing child and a missing phone can go to different teams and never wait the same time. The scheme is below and lives in one file.
- Staleness
- Built
- A timer, not a judgement. Red: 60s without acknowledgement or 2 min without a responder. Amber: 3 min and 5 min. Green: 15 min.
- Channel
- Built
- Earpiece breaks in for life safety, patterns and medical or child calls left waiting. Everything else is on the phone card. Routine issues stay in the ledger and never reach Mo.
- Length
- Built
- Spoken brief: at most three items, about thirty words, most urgent first. Card headlines: twelve words or fewer.
- Interruptions
- Built
- Two tones, then the headline. Each card interrupts once; after that it waits on the phone or for "Brief me". Mo can turn alerts off.
- Other audiences
- Partly built
- Every incident now names its team (First aid, Child-safe team, Security, Site crew, Volunteer coordinator, Info Tent), which is the basis for each team lead's own view. The filtered views themselves are designed, not built.
- Trust
- Built
- "Change or skip" on every card opens "Why you're being asked": the exact words said on air, who said them and when, what has stalled, and why the suggested person.
Response
Act with human approval
- Menu
- Built
- The AI can draft four things: a radio call, a text to named volunteers, a request for a free volunteer to move, and words for a stage MC.
- Bright lines
- Built
- It never sends a responder, stops a stage, holds entry, releases a child, calls 000 or messages anyone. It drafts; a person approves. Nothing reaches the outbox without Mo.
- Approval
- Built
- One tap, on a button that says what it does: "Send Aisha", "Text Rosa and Chris". "Change or skip" offers the alternatives: send someone else (a side-by-side compare of up to three people with what each move costs), edit the message, change the type or urgency, ask again in 2 minutes, or skip. And Mo can change his mind about anything afterwards: reclassify the type, team or urgency, send someone different, reopen something marked resolved, split or join incidents, or send a correction for a message already out. Routine work (chasing Control, sending floaters to queues, volunteer breaks, covering gaps from spare posts) runs on standing orders Mo switches on from the phone's menu at the start of the shift: each goes out after 20 seconds, logged as sent on his standing order, and the idle screen says what each open call will do next and when. Red calls, children, releases and same-incident decisions are never standing orders.
- Targeting
- Built
- Drafts name people from the roster with the right certificate and a free or non-critical post: Sam (first aid, said on air they were free) for the Lawn Stage; gate stewards only for a lost child; two floaters for the water queue.
- Evidence
- Built
- Each draft carries a one-line why. The Why view shows the source transmissions behind it.
- Timeout
- Built
- "Ask me again" hides a card for 2 minutes. A dispatch draft that is overtaken on air (Delta 2 says en route before Mo decides) retires itself and is logged. Anything still waiting comes back as a new card.
Close and write up
- Closure
- Built
- Only a confirmation on air closes an incident: fine, reunited, discharged, stood down. The child incident closes with proof ("Security 3 checked her ID").
- Handover
- Designed
- The ledger is the handover: everything unresolved, their state and their last words on air. Shift change is a filter, not a meeting.
- Report
- Designed
- End-of-day incident report drafted from the ledger, with gaps flagged for a person to fill (times, names, treatment).
- Record
- Built
- Every AI suggestion and every human decision is logged with a time and who decided. Overrides read as plain sentences next to what the AI said ("14:06 Mo changed INC-02 urgency from Amber to Red, AI said Amber"), with an overrides-only filter on the phone and in the ledger.
- Learning
- Designed
- Friday's ledger shows which posts stalled and which drafts were changed. That tunes Saturday's thresholds and wordings.
Prepare before gates open
- Pre-written words
- Built
- Dispatch, chase, lost child, child release, heat announcement, need hands and welfare break are written in advance. The AI adapts them rather than inventing wording in a crisis.
- Scenarios, triggers, roles, site facts, rehearsal
- Designed
- Written plans for heat, storm, crowding, lost child, medical and evacuation, each with a trigger, five-minute roles and pre-written words. The Sunday storm scenario is next.
Coverage
Build the roster, confirm who shows up, repair on the day
- Rostering
- Out of scope
- The brief says go narrow and we chose incidents. A small roster exists only so drafts can name real people, check certificates and see who said they are free.
- Repair on the day
- Partly built
- Where incidents and coverage meet, we built it: two no-show first aiders become an incident, and the draft proposes a first-aid-certified volunteer from a non-critical post. A struggling volunteer (Jess, in the sun since 10am) becomes a welfare incident with a drafted break and named cover.
Cross-cutting
- Connectivity
- Designed
- When 15,000 phones congest the network, radio still works. The earpiece brief is the primary channel and the phone is secondary. Outbound texts to volunteers are short enough for SMS.
- Errors
- Built
- A missed incident is worse than a false alarm for children and life safety, so those lean loud: they interrupt, and child joins always need a person. For routine issues we lean quiet. A failed model call becomes an unprocessed card, never a silent drop.
- Privacy
- Designed
- Children's descriptions go only to the stewards who need them and only during the search. Medical details stay in the ledger, not in texts. Volunteer numbers never leave the system.
- Adoption
- Designed
- Volunteers install nothing: they use radio and plain texts. Only Mo and Control use the phone view.
- Accountability
- Built
- Every outbox message says who approved it and when. The AI never appears as the sender.
- Access
- Designed
- Everything on a card can be spoken. Severity is shown in words as well as colour. Drafts use plain English a first-timer understands in ten seconds.