A venture studio for specialized AI

Build the AI company thatrewrites an industry.

25Madison incubates deep AI startups inside the industries the internet forgot — healthcare, logistics, manufacturing, construction, insurance and supply chain. We pair operators, capital and engineering firepower from day zero.

The studio

We don't fund decks. We co-found companies.

25Madison is a venture studio in NYC building AI-native operating systems for the trillion-dollar industries running the physical world. Every company starts inside the studio with a thesis, a wedge and a co-founding team.

Funded from day zero

Capital, design, recruiting and GTM are in the room before the first commit.

Operators in the room

Founders work side-by-side with industry veterans who've shipped into these markets for decades.

Real industries, real data

We unlock proprietary data and design partners so models train on what actually matters in the field.

Where we play

Horizontal AI is a feature.Vertical AI is a company.

The largest, slowest, most regulated industries are where AI creates generational outcomes. We pick the wedge, then build the moat.

HealthcareLogisticsManufacturingConstructionInsuranceSupply chain
Built inside the studio

Four companies.One thesis.

Each of these started as a whiteboard at 25 Madison. Today they're shipping into real customers, with deep AI architectures purpose-built for their industry.

Hank architecture diagram
Portfolio · 01

Hank

Voice agents for construction field teams

Hank turns every walkie-talkie, phone call and field check-in into structured project intelligence. Real-time, low-latency voice AI built for hard-hat environments — and wired straight into Procore and Autodesk Construction Cloud.

Tech under the hood
  • LiveKit + OpenAI Realtime, sub-300ms turn-taking
  • Async LLM extraction pipeline over WebSocket streams
  • Native two-way OAuth into ACC & Procore
Mason architecture diagram
Portfolio · 02

Mason

AI-native operating system for litigation

Mason ingests millions of pages of case documents into an ontology-driven knowledge graph, then runs LangGraph multi-agent workflows on top — every answer cited, every claim grounded in the record.

Tech under the hood
  • RDF/TTL ontology + Postgres semantic triples with JSONB provenance
  • LangGraph allegation hierarchy and deposition Q&A agents
  • Citation-bound LLM responses tied to source page & bbox
Ferra architecture diagram
Portfolio · 03

Ferra

Know which steel jobs to bid. Win more of the right ones.

Ferra reads raw construction docs — PDF plans, CAD, specs — and produces a fabrication-ready bill of materials, pricing and bid strategy. A multi-agent system that takes steel fabricators from ingestion to winning bid.

Tech under the hood
  • Deep-learning component identification (CNNs / transformers + OCR)
  • Takeoff & analysis agents with industry-pricing tool use
  • MAS orchestrator: planning, optimization and summary agents
Rebuild architecture diagram
Portfolio · 04

Rebuild

AI-driven restoration estimation engine

Rebuild lets restoration teams write accurate, high-dollar estimates in a fraction of the time. It fuses photos, voice intake, notes and 3D scans into precise, defensible scopes through agentic workflows.

Tech under the hood
  • Multi-modal ingestion: CV damage detection, ASR/NLU voice ops, point clouds
  • Document & visual understanding agents with industry pricing knowledge graph
  • Generative estimation model grounded in industry rules
What we look for

Engineers who want to build the company, not just the model.

01

Senior AI / ML engineers

You've shipped real ML systems — RAG, multi-agent, computer vision, voice — into production at scale.

02

Founder DNA

You want to own a problem end-to-end. You'd rather build a company than join one.

03

Domain curiosity

You're energized by gnarly, specialized industries — not chasing whichever model dropped this morning.

Join the studio

Come build the next big startupand transform an industry with AI.

Founding engineer, ML lead, technical co-founder. If any of those words make you lean in — drop your details below.

At minimum, give us a LinkedIn URL or a CV (or both).