Inspiration
Construction is one of the world's largest industries and one of its most dangerous. In India, nearly 38 workers die on site every day, and roughly 5% of every project's budget is lost to rework. What struck us was why: a single safety officer can't watch fifty cameras and every worker at once, and the most important signals — a missed curing day, a rebar that doesn't match the drawing, a worker drifting into a crane zone — either happen where no one is looking or unfold slowly over time. The information exists; nobody can process it fast enough. We wanted to give the site a sense of its own.
What it does
We built a nervous system for the construction site. Fixed CCTV gives the site a continuous, whole-site memory; Meta smart glasses give each worker hands-free eyes and an alert channel. Both feed one shared brain that runs 32 safety and quality checks — PPE, edge protection, restricted zones, concrete curing, honeycombing, bar-bending compliance, and more. Three capabilities go deeper than object detection: it remembers (counting the 14-day curing window, catching a missed pour vibration), it checks the plan (comparing live rebar against the approved bar-bending schedule), and it warns the right people — the instant it sees a hazard it speaks a warning into the nearest worker's glasses and pushes a real-time alert to the site manager, with location and photo evidence, in under two seconds.
How we built it
The core design is tiered inference, so it's cheap enough to actually deploy. Tier 1 runs lightweight YOLO-class detectors continuously on every feed as a tripwire. Tier 2 fires a vision-language model only when Tier 1 flags something worth the compute — reading drawings and judging defects. Tier 3 uses object tracking plus an event memory and RAG over the approved drawings for time-based and plan-based checks. Crucially, we sample only 2–3 frames per second per camera, never the whole stream. With $N$ cameras, cost scales like
$$C \approx N \cdot c_1 + k \cdot c_2, \quad k \ll N \cdot \text{fps}$$
where only the small number of triggered events $k$ pay the expensive $c_2$. One edge box per camera cluster runs Tier 1 locally; only events travel to a central GPU.
Challenges we ran into
The hardest problem was reasoning over time — most checks (curing, vibration) can't be answered by a single frame, so we needed reliable event state, not a classifier. Second was vision × documents: matching what a camera sees to an approved drawing is far harder than detecting a helmet. Third was cost: naïvely processing every frame from every camera is a data-centre fantasy, which forced the tiered, event-triggered architecture. Finally, deciding who to alert — mapping a hazard to the nearest worker — required fusing camera zones with worker identity.
Accomplishments that we're proud of
A single brain that spans 32 checks instead of one point-solution per problem; a closed loop that turns site-wide sensing into worker-specific action; and an architecture that's genuinely affordable and scales from one building to a whole multi-tower project without redesign.
What we learned
That the winning edge isn't a bigger model — it's memory, plan-checking, and taking action. Judges and sites alike reward systems that reason over time and do something, not chatbots wrapped around a prompt. We also learned how much cost discipline (frame sampling, tiering) determines whether an idea is deployable at all.
What's next for Construction Safety and Quality
A live Stage 2 prototype on real feeds: PPE and zone detection, the curing/vibration check, rebar-versus-drawing, and the CCTV→worker+manager alert loop — then extending the same brain to more checkpoints and integrating with Kaya's Amber as its physical-world sense.
Log in or sign up for Devpost to join the conversation.