Deterministic breakers run first
Hard rules — zoning mismatch, tax > 40% NOI, price > 1.4× comps — execute in plain Python before the LLM ever speaks. The model can enrich; it cannot upgrade past a structural problem.
Paste a Texas commercial property and an asking price. The system pulls public-record property, zoning, and tax data, runs deterministic screening rules, and returns a verdict — pursue, pursue with caution, or reject — with every claim cited to its source signal.
Four architectural rules separate a deal verdict you can stake capital against from a paragraph of generated prose.
Hard rules — zoning mismatch, tax > 40% NOI, price > 1.4× comps — execute in plain Python before the LLM ever speaks. The model can enrich; it cannot upgrade past a structural problem.
Strengths, risks, and breakers are tagged with the tool that produced them. If a claim isn't cited, it isn't in the brief. If a signal failed, it's labelled — not silently filled in.
The model returns a typed object — verdict, score, confidence, deal-breakers — that fails closed if the LLM hallucinates a field. No free-form prose to misread under pressure.
Each analysis emits an event log: tools called, inputs, outputs, cache hits, latencies. The deal brief is permalinked. You can hand it to a partner — or to your auditor.
Three actual listings run through the engine. Every verdict includes a full audit trail of the signals and rules that produced it.
Tax burden exceeds 42% of NOI — hard breaker triggered.
Confidence: 91% · 12 signals analyzed
Cap rate 50bps light, but rent growth +14% YoY offsets risk.
Confidence: 78% · 12 signals analyzed
No breakers, cap rate at market median, strong submarket comps.
Confidence: 88% · 12 signals analyzed
Sign in to access your workspace, run analyses, and view your deal history.