Forward-Deployed: Shipping Five AI Systems Zero-to-One

There's a particular kind of role that doesn't show up cleanly on a job ladder: you're handed an ambiguous business problem, no spec, no platform team, no one above the work and no one below it — and you're expected to return with something in production. That's forward-deployed engineering, and it's how I've spent the last stretch as the sole AI engineer at a live NYC commercial real estate brokerage.

Five production systems, zero-to-one, in a domain where a hallucination costs real money and a missed contact is a lost deal. This is what that work actually looks like.

Discovery is the first deliverable

The brokers didn't ask for "a skiptracing agent." They said contact data was a mess and deals were slipping. The system is downstream of a conversation. Before writing code I sat with the operators, watched the workflow, and decomposed the vague complaint into a problem an LLM pipeline could actually own — and, just as importantly, the parts it shouldn't touch.

Forward-deployed work lives or dies on this step. If you decompose the problem wrong, the cleanest code in the world ships the wrong system.

The five systems

• Skiptracer — resolves the human owner behind any NYC property through LLC chains: a NY DOS → ACRIS → IDI waterfall (BBL → ACRIS owner → entity lookup → person resolution → confidence scoring), targeting ≥95% owner and phone accuracy. The confidence scoring is the point — the team acts on the output without re-checking it.
• CompScope — a Supabase database of 9,391+ comparable NYC commercial sales with asset-type classification (multifamily, mixed-use, retail) from PLUTO and ACRIS, powering comparable-sale lookups for deal analysis.
• BuyerScope — scores and ranks buyers per asset type from ACRIS transfer history and DOB permit activity, producing a ranked top-30 list per property type for outreach (2,456 buyers scored on the Builder Score component alone).
• P1 Generator — webhook-driven MMS outreach that sends owners a one-page deal summary: CallTools → Flask → Twilio MMS, with Cloudflare R2 image storage and SendGrid.
• Call Notes Worker — transcribes and scores inbound broker calls for quality and follow-up signals: CallTools audio → Whisper → GPT-4o → structured output.

The work spans NYC public data (ACRIS, PLUTO, DOB, NY DOS), Python and Node.js, Supabase/PostgreSQL, Twilio, and the Claude and OpenAI APIs — plus ABV, an agentic valuation tool I built on the same stack and demoed to management, live at realestate-agentic.vercel.app.

Owning it means owning the boring parts

The interesting part is the agent. The job is everything around it — the ingestion that feeds it, the data model it writes to, the interface the team trusts, and the runbook for when it's wrong. When you're the only engineer, there's no one to hand the "unglamorous" 80% to. That constraint is a feature: nothing gets lost in a handoff, because there isn't one.

The bar in a high-stakes domain

In CRE, a confidently wrong answer doesn't just annoy a user — it routes a broker to the wrong owner or misprices a building. So every system ships with the same discipline I'd bring to enterprise RAG: verification before surfacing, a clear escalation path, and outputs the team can trust without auditing every time. That trust is the whole game. It's what turns a demo into a system people actually run their day on.

Forward-deployed engineering isn't a junior version of platform work. It's the opposite end: maximum ambiguity, maximum ownership, and a direct line from the problem to the person living it.