Community Session 1: Causa Claims — a Non-Coder's Five-Agent Construction-Claims Pipeline, the Gate as the Product, and What RAG Really Costs

Contract Reader (which clauses apply, notice compliance), Evidence Scout (every fact as an auditable 'evidence card': document, date, page/clause, confidence), Chronology Builder (hard rule enforced in code, not by the model: an event cannot cause a delay that started before it), Drafter (citation-backed narrative), Quantum (cost). The model proposes; the system verifies. Motivation: claims are decided on records and causation, and opposing lawyers hunt for unprovable sentences; the day-31 notice story shows the failure is a handover problem, not an AI one.

Real event: late-issued M&E/coordination drawings. Contract Reader pulls FIDIC 20.2.1, 8.1, 20.2.4 and flags notice compliance; Evidence Scout retrieves and ranks 12 items (PDF ingestion, embeddings, similarity search on pgvector); Chronology Builder on Haiku sequences the timeline with citations; the Drafter concludes the contractor is time-barred (submitted day 38 against 28); the Quantum agent applies Emden and Hudson formulas and, re-run with a hypothetical 40-day claim, shows about $2M. Each agent step displays model, latency and cost; a trust log (row hashes, JSON audit trail) was built for the hackathon judges. Opus (as heard 'Opus 4.7') for the methodology-selection reasoning, Haiku for the cheap step, an OpenAI model for ranking. Primavera P6 programme data was an input.

Praful (who has shipped similar ingestion pipelines - ~$30 just to ingest) asks for the real numbers. Syed: no granular tracking; the seven-day build ran on a ~$100/week OpenAI Codex subscription (required by the hackathon); Supabase as backend and vector store at roughly $10 per project plus ~$25 in API calls. Praful's alternative: self-host Supabase or Qdrant on a VPS and multiply instances per fixed cost. In production, documents must be re-ingested on a schedule (daily/weekly) or on events so retrieval stays current; Supabase was chosen for speed under deadline, Pinecone considered, no deep vendor comparison.

Biggest failure mode: hallucination wrapped in fluency - reviewers trust confident text more, not less. 'AI capability alone is not a product... the value is the workflow, plus the verification, plus a human at the right point.' Three layers: Generic AI (what the cohort learned - 'every one of us has the same model access'), Domain AI (30 years of claims expertise), Product AI (what the programme teaches you to build). If redoing the hackathon: one agent done well, the refusal/fluency gate first, scale later. A practising claims consultant found three gaps in ten minutes; he is rebuilding from the ground up for launch within a month or two.

Orchestration was done by prompting ('loop engineering', his term) into standalone Python/LangChain scripts, following his own domain roadmap (contract -> narrative -> cost), not OpenClaw, Hermes or Codex's native agents. He recommends the community learn LangChain and LangGraph for chaining and suggests Outskill teach them; Shivani takes it as the next member-taught topic. Most build time went to iteratively verifying outputs rather than learning new techniques. Side threads: Rahul's unresolved question on lip-synced Indian-language AI video (Praful has a Hindi example in the AIC Mastermind group); community-session format may merge with Outskill's 'Forge' meetings; Cody's Web-MCP is in a separate OpenAI contest.