← Work

1st-place hackathon prototype onto the core roadmap: AI search for events no human ever tagged

Lytx · Senior Product Manager · 2026

1st place hackathon → funded roadmap itemFinds untagged events in ~3sOne enrichment pipeline, reused across video surfaces

Problem

Fleets capture millions of video events a month, but finding a clip means already knowing its metadata — vehicle, date, event type. A manager can describe an incident ("truck swerving on a wet highway at night") yet has no way to search by what actually happened in the footage. The naive fix — more filters and tags — fails at the root: filtering only surfaces categories that already exist in the taxonomy, when the point is discovering events nobody labeled. The context is in every frame; it just isn't searchable.

Research

The reframe: treat it as information retrieval, not tagging — turn each event's video into searchable text, then match queries by meaning. Productizing exposed the real enemy: a lossy video-to-text-to-search chain, where anything the model doesn't describe becomes unfindable. That drove the rule — human review stays ground truth; search narrows while existing filters slice — so the AI never overrides the deterministic path.

Solution

Impact

The prototype won 1st place at the hackathon and earned a CEO-review greenlight onto the core roadmap — that prototype-to-funded-initiative jump is the result. At demo scale it returned ranked matches in ~3 seconds for events no human had tagged. Built as an event-enrichment pipeline, not a single search box, its summaries, index, and query log became the substrate later intelligent-video work builds on.

Reflection

The winning insight was small: you cannot filter your way to a label nobody applied — so stop tagging and start understanding the footage. The harder lesson came after the win — the demo is the easy 20%: a two-day prototype proves the idea, but the real work is hardening the lossy pipeline, cutting cost by an order of magnitude, and clearing legal. A prototype earns the greenlight; it doesn't shorten the road.

Stack

Vision-language modelLLM summarizationVector + keyword searchHuman-in-the-loop grounding
Previous$1M+/year mapping bill cut ~90%: an in-house OpenStreetMap geocoding service that replaced Google MapsNextReconciling an acquired platform's admin with the existing product so each side's customers get best of both worlds