OMO — Off Market Office

Hiring data in, square meters out.

An AI agent for Paris commercial real-estate brokers that spots companies about to outgrow their office and companies about to release one, matches them live on a map, and drafts the outreach for both sides — before either company calls a broker.

---

🎯 The problem

Commercial real estate in Paris runs on timing. The best deals are off-market — a scale-up about to run out of desks, a company quietly downsizing, a firm in redressement judiciaire about to vacate a floor. By the time these hit a broker's inbox, the deal is gone.

The signals that predict them already exist — hiring surges, exec hires, layoffs, insolvency filings — they're just scattered across sources no broker watches in real time.

💡 What OMO does

OMO watches two populations of Paris companies on one map and matches them before either side reaches out:

	🟣 Outgrowers (demand)	🔴 Releasers (supply)
Who	Hiring faster than their office can hold	Shrinking, insolvent, relocating, going remote
Signals	Hiring surge, new exec hire, job change	Insolvency filing, layoffs, champion exodus, office listing
Agent output	Months-to-breach, urgency, decision-maker, outreach	Available m², availability date, decision-maker, outreach

The agent scores each company, enriches the decision-maker, matches compatible pairs, and drafts a multi-channel outreach cadence for both — then hands the approved list to an AI sales agent. A match renders as an animated arc between the two pins on the map.

The human stays in the loop the whole way: the agent drafts and recommends; the broker approves.

🎬 The hero demo (press S)

One keystroke fires the full scripted sequence — deterministic, offline-safe, ~12 seconds:

Step	What happens
⚡ Signal	A hiring spike lands on Cartesia Labs (11e)
🧠 Math	33 people · 4.5 hires/mo · 40 desks → breach in weeks
📊 Score	Claude rates urgency 94/100 with a grounded rationale
🔧 Enrich	FullEnrich waterfall finds the Head of Workplace
🤝 Match	Matcher finds Atelier Numérique (3e) releasing 520 m² → score 91
✉️ Draft	Claude writes a multi-channel cadence for both sides

The map draws a violet→coral arc between the two pins, a toast fires, and one click opens the deal.

🖼️ A look around

• Map view — every Paris company as a pin (violet = needs space, coral = releasing), size ∝ urgency, pulsing when hot. Hover for a mini-card; click for the full deal panel. A live agent-activity popup streams the reasoning.
• Table view — a sortable "Prospects" grid with side / office-m² / urgency filters, checkboxes to build a contact list, and saved lists.
• Detail panel — urgency ring, capacity math, signal timeline, FullEnrich contact, ranked matches with Claude rationale, and a multi-channel cadence (email · LinkedIn · call) with an A/B subject test, in English or French.

🧠 Why this is an agent, not a single call

OMO isn't one retrieve-then-answer prompt. It's a pipeline of judgment and action where deterministic code and Claude each do what they're best at:

• Code owns the numbers — desk capacity, months-to-breach, needed m², and the size/timing/location match fit are computed in spacemath.ts / matcher.ts. They're reproducible and never hallucinated.
• Claude owns judgment & language — it weighs signals into an urgency score with a rationale, writes the match rationale, and plans the outreach cadence (with an A/B subject choice it reasons about).
• Adapters own the outside world — every external service sits behind an interface with a mock default, so the demo runs fully offline and any real provider is a drop-in.

🔬 How the agent loop works

flowchart LR
  S[⚡ Signal in] --> M[🧠 Deterministic math<br/>desks · breach · m²]
  M --> C[📊 Claude<br/>urgency score + rationale]
  C --> E[🔧 FullEnrich<br/>decision-maker contact]
  E --> X[🤝 Matcher + Claude<br/>size · timing · location fit]
  X --> D[✉️ Claude<br/>multi-channel cadence]
  D --> A[🚀 Max<br/>hand off the list]

Loading

Signals arrive from the seed, a live POST /api/simulate/signal, or the real feeds. Each Claude call streams a one-line summary to the agent console over Server-Sent Events, so you literally watch the agent think. Any Claude failure emits an error line and falls back to canned text — the UI never blanks.

🌐 Real data & integrations

OMO orchestrates four real external sources plus Claude — the app runs offline on synthetic data by default, and each provider goes live when its key is present.

Source	Role	How it's used
🧠 Claude (Anthropic, Sonnet 4.6)	Judgment & language	Urgency scoring, match rationale, multi-channel outreach cadence — JSON-mode calls, low temp for scoring, higher for copy
📡 Sillage (v2 API)	Demand signals	Live feed of the team's tracked accounts and their hiring / job-change / exec signals
🔔 BODACC (OpenDataSoft)	Supply signals	France's official insolvency registry — real Paris procédures collectives imported onto the map as distressed sellers, each scored live by Claude
🔧 FullEnrich (v2 waterfall)	Contacts	Finds the decision-maker's email / phone / LinkedIn for a company
🚀 Max (Digital Crew, REST v1)	Outreach	"Contact via Max" pushes the approved list to Max's AI sales agent as a real prospect list, ready to run a campaign

Real companies from BODACC are clearly badged LIVE · BODACC (green ring on the map) and kept separate from the synthetic demo set. Contacts are enriched only with consent.

🧮 The math (reproducible, in code)

Everything runs on the French office norm of ~10 m² per person.

capacityDesks   = round(officeSqm / 10)
monthsToBreach  = (capacityDesks − headcount) / hiresPerMonth        # null = over capacity
neededSqm       = round(headcount + openRoles × 0.7) × 10            # ~70% of open roles fill

The matcher scores a demand↔supply pair on three deterministic axes, then asks Claude for the rationale:

sqmFit      = 100 − min(100, |availableSqm − neededSqm| / neededSqm × 100)
timingFit   = 100 if space frees up before the breach window, decays after
locationFit = 100 − 12 × (arrondissement-distance step, haversine-derived), floor 20
score       = 0.45·sqmFit + 0.35·timingFit + 0.20·locationFit

🧰 Tech stack

Layer	Technology
Frontend	React 18 · TypeScript · Vite · Tailwind · react-leaflet (CARTO Positron tiles) · zustand · lucide-react
Backend	Node 20 · Express · TypeScript · in-memory store → server/data/db.json · Server-Sent Events
AI	Anthropic @anthropic-ai/sdk — claude-sonnet-4-6 (JSON-mode scoring/matching, prose cadences)
Adapters	SignalProvider (Sillage · BODACC · mock) · EnrichmentProvider (FullEnrich · mock) · OutreachProvider (Max · mock) · LLM (Anthropic)

🚀 Run it locally

cp .env.example .env       # add ANTHROPIC_API_KEY (see below)
npm install
npm run dev                # server :3001 + client :5173 (Vite proxies /api)

Open http://localhost:5173.

• .env holds secrets and is gitignored. Only ANTHROPIC_API_KEY is needed for the full agent demo.
• PROVIDERS=mock (default) keeps every external call mocked and fully offline-safe. Without an Anthropic key the app still runs — scoring and drafts fall back to deterministic text.
• Enable real providers by adding keys and PROVIDERS=…,sillage,fullenrich,max:
  • Sillage — SILLAGE_API_KEY (auto-detected).
  • FullEnrich — PROVIDERS=fullenrich + FULLENRICH_API_KEY (enriches a known, consented contact).
  • Max — PROVIDERS=max + DIGITALCREW_API_TOKEN (a max_live_ key). Base URL defaults to https://max.digitalcrew.tech. Set MAX_DRAFT_ONLY=1 to build the campaign without auto-sending.
  • BODACC — always available, no key.

🎛️ Demo controls

Key / control	Action
S	Fire the hero sequence (auto-switches to the map)
A	Score every company with Claude
R	Reset to the synthetic seed (confirm)
📡 top bar	Live Sillage feed (tracked accounts & signals)
🔔 top bar	Live BODACC insolvencies + "Add to map as sellers"

📁 Project structure

off-market-office/
  client/                 React + Vite + Tailwind + Leaflet
    src/components/        MapView · TableView · DetailPanel · MatchArc ·
                          AgentConsole · ContactList · Sillage/Bodacc feeds …
  server/
    src/
      pipeline.ts          ingest → math → score → enrich → match → draft
      matcher.ts           deterministic sqm / timing / location fit
      spacemath.ts         capacity math (desks · breach · needed m²)
      prompts.ts           all Claude prompts
      llm.ts               Anthropic wrapper (JSON-mode helper, SSE events)
      providers/
        signals/           mock · sillage · bodacc
        enrichment/        mock · fullenrich
        outreach/          mock · max
    data/db.json           regenerated by the seed
  .mcp.json                Sillage MCP server config (agent-side)

🔒 Responsible use

• All demo companies, people, and contacts are synthetic and fictional. Real BODACC records (public legal notices) are clearly separated and badged LIVE.
• Human-in-the-loop: the agent drafts and recommends; the broker approves. Nothing is sent without a connected account and an explicit action.
• Contacts are enriched only with consent; the map footer notes the synthetic data.

🏆 Built at the Agentic GTM Hackathon

Station F, Paris — Anthropic × FullEnrich × Sillage × Digital Crew. Deterministic math in code, Claude for judgment and language, real external data on both sides of the market.

OMO — hiring data in, square meters out.