LargeKite Research Lab
Experiments in AI-augmented real estate investing
Each entry below is real infrastructure shipping on the platform — not a roadmap of aspirations. Where an experiment surfaces in product, the link goes to the live feature; where it's still in R&D, we describe the design and the open questions.
Multi-Agent Investment Committee Pipeline
Five specialist agents — Underwriting, Market Research, Risk, Operations, and IC Synthesis — collaborate in parallel to produce institutional-grade deal analysis.
Specialists run concurrently against the same DealInput; the IC agent runs sequentially, consuming the other four outputs to synthesize the final thesis, strengths, weaknesses, and recommendation. Each agent enforces a strict JSON contract — model output is validated and falls back to a typed stub when the API key is missing, so the orchestrator degrades gracefully rather than crashing.
Structured OM Extraction
Broker Offering Memorandums become typed JSON: purchase price, NOI, occupancy, unit mix, expense breakdown, broker assumptions, and red flags.
The extraction agent receives parsed PDF text and is constrained by a strict schema. Numbers must be present in the document or returned as null — the model is explicitly forbidden from interpolating or guessing. Red flags are surfaced when the model detects aggressive broker assumptions (pro forma vacancy below trailing, missing reserves, excluded management fees, etc.). On well-formatted OMs, headline-financial accuracy exceeds 95%; the failures are loud rather than silent.
Market Sentiment Engine
Composite market sentiment derived from 20-day momentum, VIX positioning, and 50-day-vs-200-day trend regime — refreshed on each request.
The engine ingests real Yahoo Finance series for the S&P 500 and VIX, computes moving averages, and blends a short-window momentum signal with a volatility-adjusted trend indicator. The headline pill on the homepage only shows "Hot" when the 50d trend is above the 200d trend AND short-window momentum is positive — preventing single-day spikes from being misread as regime changes.
Triple-Shock Stress Test
Three concurrent shocks applied to base cash flow — vacancy doubles, rents 10% below estimate, and rates +100 bps — to test whether a deal survives the worst plausible scenario simultaneously.
Rather than running one-at-a-time sensitivity tables (which understate joint risk), the stress test applies all three shocks at once and reports the surviving cash flow. The output drives both the in-product warning band and the IC memo stress section. Deals that lose more than 60% of base cash flow trigger an automatic "Conditional" cap on the IC recommendation regardless of other inputs.
Property Scoring Model
A 0–100 composite score combining cap rate, cash-on-cash return, cash flow, and strategy fit — calibrated against 50-state tax rates and submarket-specific rent norms.
Every property surfaced by the finder receives a structured score using documented weights. The model is intentionally transparent — each component contributes a known number of points, and the breakdown is exposed in the UI so investors can see why a deal scored what it scored. Strategy fit (rental, BRRRR, house-hack, appreciation, STR) re-weights the components dynamically based on the user's declared strategy.
Climate Risk Overlay
Layered physical-risk signals — flood, wildfire, hurricane exposure — sourced from public data and applied to property-level analysis as a soft warning band.
Climate risk does not currently affect the headline investment score, by design — we want investors to see the risk and decide for themselves. The overlay flags properties in elevated-risk zones with the underlying source (FEMA, USFS) cited inline. Next iteration will add insurance-cost differential estimates.
Secondary-Market Scoring Framework
Five-variable framework for evaluating secondary US markets: household formation, employment diversification, supply discipline, affordability spread, and regulatory stability.
Currently a written framework applied manually in the IC memo Market section. Roadmap: automate the scoring into a queryable submarket table, refreshed quarterly, exposed to the Market Intelligence agent as structured grounding. Will eventually power a "comparable markets" view that ranks similar submarkets when analyzing a target.
Open Questions
What we're trying to figure out next
- →Can a "comparable deals" memory layer materially improve IC committee decision quality on its eighth-year deals?
- →What is the right level of automation for the "why we should pass" pre-mortem section — and does AI hurt or help analyst judgment?
- →How do we surface broker assumption-stack quality as a first-class metric that committees can calibrate against over time?
- →Can submarket-specific supply pipeline data be ingested from CoStar / Yardi / public permit datasets reliably enough to ground the Market agent?
If any of these are interesting to you — as an operator, analyst, or researcher — we'd like to hear from you. Get in touch.