About this Project
A working multi-agent AI investment platform,
built from first principles.
LargeKite Capital Intelligence is the platform an institutional real estate investment committee would build if they started over today: deterministic financial math at the core, AI agents for everything else, persistent memory across deals, and an output that matches what a human IC analyst produces — not a chatbot, not a calculator, not a spreadsheet wrapper.
Why this platform exists
Real estate investing is two problems pretending to be one. The first problem is numerical — what is the cap rate, what is the cash flow, what is the debt service coverage ratio. This problem is solved. Any spreadsheet can do it.
The second problem is judgment — is this market structurally strong, are these rent assumptions defensible, what is the broker not telling me, what would an experienced investment committee actually argue about. This is where retail tools collapse and institutional platforms cost $3,000+ per month per seat.
LargeKite was built to collapse that gap. The same multi-agent pipeline a $50B PE firm would commission internally, exposed as a working product. Not as a demo. Not as a white paper. As something you can run on a real broker memo and get an institutional investment-committee analysis in under a minute.
Four design principles
The decisions that shape every feature on the platform.
Real institutional output, not a demo
Every deal that runs through the analyzer produces six discrete agent outputs — underwriting, market research, risk, operations, IC discussion, and IC synthesis — that an institutional investment committee would actually consume. No fake numbers, no scripted answers. The same JSON contracts the IC report renders from are the ones a human analyst would produce.
Deterministic where it matters, AI where it helps
Cap rates, DSCRs, cash-on-cash returns, and tax math are calculated deterministically — same input, same output, every time. AI is used for what humans are slow at: extracting structured data from OM PDFs, synthesizing multi-source market narratives, surfacing hidden risks from rent rolls, and writing investment-committee-quality summaries. The split is intentional: numbers are pure functions, judgment is augmented.
Memory, not stateless tools
Every analysis is remembered. Re-analyze a property six months later and the platform surfaces what changed — rent trajectory, market shift, deal-level deltas. Most retail tools forget the moment you close the tab. Institutional investors track deals over months and years; the platform was designed to match that workflow.
Production-grade, not portfolio-grade
Server-rendered Next.js, typed end-to-end, full SEO + Open Graph, RSS + JSON feeds, sitemap, structured data, accessibility-aware components, mobile-first, installable as a PWA on iOS. The platform ships, not just demos.
How the AI pipeline works
The architecture in five layers — each one inspectable, each one with a strict contract.
Underwriting, Market, Risk, and Operations agents run concurrently against the same DealInput. Each enforces a strict JSON contract — invalid output triggers a typed fallback so the orchestrator degrades gracefully rather than crashing.
After specialists complete, a discussion agent simulates how an investment committee would actually debate the deal — surfacing disagreement, weighting concerns, and producing a committee-style transcript.
The synthesis agent consumes all upstream outputs and produces the final thesis, strengths, weaknesses, recommendation, and key conditions. This is what renders as the downloadable investment memo.
A persistent store of prior analyses. When you analyze a deal you have seen before, the platform surfaces what changed and what the prior thesis got right or wrong.
Cap rate, DSCR, cash-on-cash, IRR, debt service, recast NOI, hidden-risk pattern matching, and threshold-based scoring. Pure functions, fully testable, no AI involvement.
What has shipped
A real platform, not a prototype
Application Context
This project is part of an application to the University of Chicago Booth School of Business MBA program.
Booth's admissions process values evidence of disciplined thinking, the ability to ship under uncertainty, and intellectual honesty about what you learned along the way. This platform is a working artifact of all three.
The product is not a hypothetical. It is a multi-agent AI investment platform that produces institutional-quality investment-committee analysis on real broker offering memorandums. The codebase is production-grade — typed end-to-end in TypeScript, server-rendered, SEO-complete, PWA-installable, with deterministic financial math separated from the AI surface area so every number is reproducible.
The architecture decisions, the trade-offs that did not work, and the build cadence are all documented publicly in the Field Notes and Research Lab. The institutional thinking — what makes a strong secondary market, how to read a rent roll for hidden risks, how AI changes underwriting cycle time — lives in the Research section.
The goal of an MBA at Booth is not to learn what to think. It is to learn the analytical frameworks for thinking better, faster, and more honestly than you could alone. This platform was built in that spirit: every claim is traceable to its inputs, every output is inspectable, and every decision has a reason that is written down.
About the builder
LargeKite is a one-person research and build effort. Every line of code, every research piece, every design decision in the platform was made by one person working out loud — the field notes, the changelog, and the architecture writeups are the receipts.
A fuller personal statement — background, motivation, what this project taught me about the gap between retail real estate tools and institutional ones, and what I want to do at Booth and after — is in progress. In the meantime, the work itself speaks: read the field notes for the build journey, the research articles for the analytical voice, and the research lab for the architectural decisions.
For specific questions or to discuss the project further, please reach out via the contact form.
What is next
The platform will keep shipping. The next priorities are saved-search alerts so investors can track new listings without re-running the finder, a deal pipeline view that mirrors how investors actually move properties from watching to closed, and broader data source coverage beyond a single MLS API.
Longer term, the institutional thinking gets sharper: more markets, more case studies of deals analyzed end-to-end, a portfolio-level view that combines individual deal analysis into portfolio-level risk and return modeling.
The build is logged in real time in the changelog. Subscribe to the RSS feed to see new research as it ships.
Related
Keep reading
The story above is the high-level. These pages are the depth — read in any order.
Platform by the numbers
200+ commits, 60+ routes, 13 markets, 5 research articles — the build at a glance.
Research Lab
How the multi-agent AI pipeline is wired, with file paths and live experiment status.
Long-form research
Five articles in institutional voice on underwriting, market selection, and AI workflows.
Case studies
Three deals taken end-to-end — Tampa multifamily, Phoenix market review, investment memo template.
Field notes
Five short pieces on what shipped this week, what surprised me, and how the patterns evolved.
Walk the demo
A guided tour of the platform from OM upload through IC synthesis on a sample deal.