Proof I build, not just advise
Prototypes & Builds
For most of software history, building a product meant deep teams and long cycles. Vibe coding collapses the cost of a prototype — but it doesn't change what makes a product good: a real problem, a sound domain model, controls you can trust, and distribution. Everything here is built the same way I ship in production — foundation first (schema, contracts, security), then AI on top — because coding agents multiply whatever foundation you give them.
Working builds
Four products, built end-to-end
AI-Native Property Claims Platform
A full property-claims lifecycle product — not a chatbot demo — spanning customer intake (form, voice, phone), a role- and region-aware Claims Desktop, a propose-only staff AI assistant, an explainable fraud engine (rules + LLM blend with SIU dispositions and payout holds), payments (Stripe Connect), and AI quality operations (health, drift, golden + red-team evals).
Stack: Next.js/Vercel · Claims-MCP on Render · n8n Cloud · ElevenLabs · Langfuse · Supabase (RLS)
The write-upFACIA — FAst Cat Impact Analytics
Turns a weather signal into insured-impact intelligence in minutes, not hours of spreadsheet work. A carrier ingests a portfolio; the system monitors wind/hail/rainfall thresholds against gridded NOAA data, detects material events, synthesizes exposure footprints with confidence scoring, classifies policies as directly or adjacently impacted, and hands operators a live map + exportable impact package — target under 30 minutes from threshold breach to report.
Stack: React/Vite/Mapbox · FastAPI · CrewAI · n8n · Supabase Postgres + PostGIS + Realtime · NOAA MRMS/URMA · Anthropic Claude (narrative only, numbers from the DB)
The write-upWorkers-Comp FNOL Copilot
Smart injury intake & triage that turns fragmented Workers-Comp First-Notice-of-Loss into one auditable incident record. A mobile-first PWA guides injured workers step-by-step with optional empathetic AI batch chat; supervisors contribute via secure invite links; HR gets completeness scoring, rule-based triage tiers, and FNOL submission readiness before a payload leaves the building.
Stack: Next.js PWA · Supabase (Postgres, Auth, RLS, Edge Functions, Storage) · shared TypeScript domain layer · 80+ automated tests
The write-upRestaurant Voice Ordering & Operations Copilot
A voice- and chat-first ordering copilot (piloted for a real Indian kitchen) that feels natural to guests but treats the database — not the model — as the source of truth. Handles natural multi-turn ordering, safety-critical allergen classification (contains / may contain / cross-contact / unknown, never claiming medical safety), full read-back and explicit confirmation before a kitchen ticket, and a staff ops dashboard. A deliberately out-of-domain "range" build that shows breadth and safety-first agent design.
Stack: Convai reasoning · deterministic server tools · fail-closed webhook auth · 114 automated tests
The write-upLive demos and screenshots are on their way.
01
How to Build Well (in the vibe-coding era)
Foundation beats flash: schema, RPCs, and invariants before agent demos. Half the product is contracts (auth, webhooks, tool permissions, RLS) — not the model.
Ground agents in tools, not memory. Split conversation from commitment: the LLM handles dialogue; the backend enforces state transitions. Ship trust as a feature: observability, golden + red-team evals, and drift monitoring go in with the agents, not on a later slide.
02
Anti-Patterns & Failures Guide
Where vibe-coded builds go wrong: no problem definition (a demo that solves nothing); letting the model "remember" facts that must come from data (hallucination); skipping evaluation and governance; treating multi-tenant security as a final hardening step instead of the first migration; duplicating business logic across edge and domain code; and confusing a plausible UI with a correct one.
Honest, specific, and hard-won.