Fondok AI

AI-powered hotel acquisition underwriting. From offering memorandum to investment committee memo in 17 minutes.

Fondok is decision infrastructure for institutional hotel investors. It ingests broker materials (OMs, T-12s, STR reports, term sheets), runs them through a multi-agent extraction + reasoning pipeline, and produces a fully-cited IC memo with live IRR, variance detection against the broker case, and exportable Excel / PDF / PPTX deliverables.

Live demo: https://fondok-app.vercel.app/
Inspiration: https://fondokai.lovable.app/
Worker API: https://fondok-worker-production.up.railway.app/

---

What it does

1. Ingest. Drop a folder of deal documents — Fondok classifies each (T-12, OM, STR, term sheet, …) and routes it to the right extractor.
2. Extract. Per-document field extraction with confidence scores and citations back to the source page.
3. Underwrite. Three engines (returns, variance, comp set) compute IRR / EM / CoC and flag where the broker case diverges from the underwritten case.
4. Synthesize. Streaming IC memo generation in 6 sections, every claim cited.
5. Export. Excel model, PDF memo, PowerPoint deck.

---

Architecture

┌──────────────────────┐         ┌────────────────────────────────────────┐
│ Next.js 14 web app   │  HTTPS  │ FastAPI worker (Railway)               │
│ apps/web (Vercel)    │◄───────►│ apps/worker                            │
│  - Landing + dashboard│         │  - /deals  /documents  /memo/stream    │
│  - Live IRR engine    │         │  - LangGraph multi-agent orchestration │
│  - SSE memo viewer    │         │  - Anthropic SDK (cached prompts)      │
└──────────────────────┘         │  - Postgres + Redis                    │
                                 └──────────────┬─────────────────────────┘
                                                │
                              ┌─────────────────┼─────────────────┐
                              ▼                 ▼                 ▼
                       ┌────────────┐    ┌────────────┐    ┌────────────┐
                       │ Claude     │    │ Claude     │    │ Claude     │
                       │ Haiku 4.5  │    │ Sonnet 4.6 │    │ Opus 4.7   │
                       │ classifier │    │ extractor  │    │ memo synth │
                       └────────────┘    └────────────┘    └────────────┘

Routing logic chooses the cheapest model that meets the task's bar; prompt caching cuts repeated-context cost on extraction reruns and memo regeneration.

---

Quick start

# install
npm install                       # workspace root
cd apps/worker && uv sync         # python deps

# run the web app
cd apps/web
NEXT_PUBLIC_WORKER_URL=https://fondok-worker-production.up.railway.app npm run dev
# unset NEXT_PUBLIC_WORKER_URL to render with mock data

# run the worker locally (see apps/worker/README.md for full setup)
cd apps/worker
uv run uvicorn app.main:app --reload --port 8000

---

Tech

• Web: Next.js 14 (App Router), TypeScript, Tailwind CSS, Clerk auth (optional)
• Worker: FastAPI, LangGraph, Pydantic, Anthropic SDK, Postgres, Redis
• Models: Claude Haiku 4.5 (classification) · Sonnet 4.6 (extraction) · Opus 4.7 (memo)
• Infra: Vercel (web) · Railway (worker + Postgres + Redis)

---

Phase status

Phase	Scope	Status
1	Mock-data UI shell, dashboard, projects list	done
2	Worker API contracts, deal CRUD, doc upload	done
3	Live worker integration, classification, extraction	done
4	Memo SSE streaming, citations, exports	done
5	Variance engine, comp sets, scenarios	done
6	Landing page, favicon, demo polish	in flight

---

Documentation

• docs/fondok-ai-walkthrough.md — product walkthrough
• apps/web/README.md — web app smoke test + env vars
• apps/worker/ — FastAPI source (see app/api/ for endpoint contracts)
• evals/ — golden-set evals + last run results
• DEPLOY.md — deployment runbook (Vercel + Railway)

---

Diagnostics

A non-linked diagnostics page is exposed at /diag showing worker URL, /health, deal count, recent agent costs, and an in-browser smoke test runner.