Pulse.ai

Inspiration

Construction runs late almost everywhere — 98% of megaprojects run over budget or behind schedule (McKinsey), and 70% of projects report supply-chain disruptions as a major concern. When we dug into why, the pattern was always the same: materials get ordered, and then the chaos begins.

A material's real status is scattered across purchase orders, vendor WhatsApp messages, ERP systems, delivery challans, and site photos — and none of it is connected to the schedule task that is waiting on that material. So teams find out a material is late only when the crew is already standing idle, by which point the delay has already cascaded into the schedule and the budget. Site managers told us they lose up to 35% of their time just chasing information.

Every tool we saw only logged the past. None of them predicted whether a material would actually arrive on time. Track 2 asks exactly that question — so we set out to give the construction supply chain a brain.

What it does

Pulse turns fragmented material updates into days-ahead warnings. It follows a simple loop:

Ingest → Graph → Predict → Act

Ingest — pulls material status from POs, delivery challans, vendor messages, and delivery photos. No new manual data entry — humans confirm, they don't re-type. Graph — links every material to the schedule task that depends on it, weighted by critical-path urgency. Predict — forecasts each material's expected arrival and compares it to its need-by date, producing a live red / amber / green risk score ranked by schedule impact. Act — when a material turns red, an AI agent drafts an expedite email to the vendor, surfaces alternate vendors, alerts the responsible owner, and proposes a schedule re-sequence.

Humans see and steer the brain through three interfaces: the Risk Radar (materials ranked by schedule risk), a Copilot (ask anything, end-to-end — "what's blocking the 3rd-floor pour?"), and proactive alerts with drafted actions.

How we built it

Pulse is built as three technical pillars:

Perception — Document Intelligence + Vision-Language Models. OCR and a VLM extract structured status events from messy, unstructured inputs. A vision model even reads delivery photos at the gate, checking delivered quantity and flagging damage automatically. Memory — a Material→Schedule Risk Graph. A knowledge graph connects materials, vendors, deliveries, and schedule tasks, with edges weighted by float and critical-path status. GraphRAG rides on this graph to power the copilot with grounded, connected answers. Foresight & Action — Forecasting + Agentic Systems. A forecasting engine estimates arrival from each material's status trail and the vendor's lead-time history, then scores risk:

Risk=P(arrival>need-by)×task criticality\text{Risk} = P(\text{arrival} > \text{need-by}) \times \text{task criticality}Risk=P(arrival>need-by)×task criticality When risk crosses the line, a tool-using agent takes concrete action rather than just notifying.

Stack: the app runs on Next.js + Supabase + Vercel, with the Anthropic API (Claude) powering extraction, the copilot, and the agent, a graph database for the material–schedule model, and Python for the computer-vision and forecasting layers.

For the hackathon we scoped tight — one project, one high-value long-lead material class, and three ingestion types (a PO, a vendor dispatch update, and a delivery photo) — so we could prove the full loop end to end.

Challenges we ran into

Every vendor speaks a different language. Status arrives as PDFs, WhatsApp texts, and emails; normalizing all of it into one consistent "material status event" schema was the hardest engineering problem.

Reading imperfect site photos. Getting reliable quantity and damage signals from real, badly-lit delivery photos took iteration.

Connecting status to the right task. Modeling the schedule as a dependency graph and computing critical-path impact — so we flag the one delay that matters, not all of them equally — was subtle.

Cold-start forecasting. With sparse delivery history, we started with transparent heuristics plus vendor lead-time priors, designed to sharpen into a learned model as data accumulates.

Keeping the agent safe. The agent drafts actions for a human to approve — it never auto-sends — so it assists without overstepping.

Accomplishments that we're proud of

We built a working prototype that runs the entire loop on a real material class — and in our demo, Pulse caught a delay a full week before the crew would have.

What we learned

The hard part isn't the model — it's the fragmentation and latency of information. The status already exists; the value is connecting it to the schedule and closing the gap before the critical path.

Prediction and action beat another dashboard or chatbot. The moment that changes everything is catching a delay before anyone would have noticed.

Integration-first wins. If a tool demands new manual data entry, a real site will never adopt it.

Domain modeling is the differentiator. Critical path and float are what turn raw tracking into genuine intelligence.