AI deal screening for a New York real-estate investor

Problem

Analysts screened loan and equity deals by hand: pulling market data from a dozen sources, benchmarking against RXR's target markets and writing a memo. Days per deal, and every number had to be traceable for the investment committee.

What I built

• Address in → ten parallel skills on Trigger.dev (market fundamentals, demand drivers, demographics, affordability, transit, crime, exogenous risk, sponsor profile, capital markets, regulatory), each returning a versioned envelope with citations, methodology, data vintage and cost.
• Progressive-disclosure dashboard: executive summary → category scores on a visible 0–10 rubric → metric detail with a click-through provenance trail on every number; outlier flags benchmarked against RXR's 12 target markets.
• Data layer on Postgres (Neon, then AWS Aurora, Drizzle): ZIP-to-submarket crosswalk, Zillow rent index, Census ACS, market benchmarks for ~200 geographies, monthly refresh job.
• Document lanes: investment-committee memo review (LlamaParse + S3), offering-memorandum upload → screening memo; asset tracking with monthly re-runs and threshold alerts; Trepp comps and sponsor loan book; FRED capital-markets series; AI cost tracking per run.
• Anti-hallucination design: grounded search → structured extraction → only real citations shown; explicit missing-data markers instead of guesses.
• Shipped to RXR's AWS (Docker on ECS Fargate, CI/CD to ECR); skills on Trigger.dev cloud. Ran a data-provider evaluation (Green Street, CoStar, Trepp and others).

Result

A screening that took an analyst days now returns in about 1–2 minutes across ~11 categories, with every figure traceable to its source. Repositories handed over to RXR's GitHub organisation in June 2026 after a structured client feedback loop.

Screens

Stack

• Next.js 15
• React 19
• TypeScript
• Tailwind 4
• Drizzle
• Postgres
• Trigger.dev v4
• Anthropic SDK
• OpenAI
• Gemini grounded search
• AWS S3 / ECS
• LlamaParse