The thesis
The operating model behind both motions.
AI Foundry Ventures runs in two motions. Senior engineers deploy into a business to find the workflow actually worth building; an agent workforce then builds it in weeks on a platform we have already hardened elsewhere. Some of what comes out is a partner’s system. Some of it becomes a company the studio holds founding equity in.
01
Production sprints, not prototypes.
Every venture AIFV builds ships in 2–4 week sprints to a working production environment. We don't deliver Figma files and slide decks; we deliver running code, deployed, observable, and handover-ready. The sprint cadence forces decisions early, exposes integration risks before they compound, and means a venture has paying customers — or a measurable reason it doesn't — within a quarter, not a year.
Proof
CortexData shipped its 5-module lending OS, 8 loan products, and full RBI compliance map in successive sprints. FlikVault went from blank repo to live academies on two continents the same way. Hundi’s settlement engine, compliance pipeline, and three-jurisdiction entity structure were assembled module-by-module with the same cadence.
02
Regulators in the room from day one.
Compliance is the substrate, not the wrapper. RBI Master Directions for CortexData. ADGM/FSRA Innovation Test Licence and RBI observer-node access for Hundi. Federation reporting standards (ECB, BCCI, ICC) for FlikVault. We design the data model around the audit trail the regulator will eventually ask for, not the other way around — which is why we ship every regulatory return built-in instead of as an integration project.
Proof
CortexData ships 9+ RBI returns out of the box: KFS, IRACP, PSL Form A, DSB, OSS-3, Co-Lending, Securitisation, Gold Loan, and audit-immutable retention. Hundi embeds FATF Travel Rule with IVMS101 at the protocol level. FlikVault produces federation-grade attendance and progress audits.
03
AI in the core, not bolted on.
Pose estimation in SKrutin. ML decisioning with full feature attribution in CortexData. Claude-written coaching notes in SKrutin. Travel-Rule-with-ZK compliance proofs in Hundi. The AI isn't a feature page on the marketing site; it's how the product fundamentally works. Which means every venture is built on infrastructure that benefits from each new generation of frontier models, not stranded by them.
Proof
CortexData’s 5-model fraud ensemble + calibrated PD scorecard returns per-decision feature attribution. SKrutin runs RIFE interpolation, YOLO/RF-DETR ball tracking, and pose-based biomechanics scoring on iPad. Claude turns every technical analysis into an age-appropriate coaching note.
04
Two motions, one loop.
A foundry answers whether a thing can be built well. It is the wrong question to start with. The one that decides whether anything gets paid for is whether this particular business should want it built at all — and that can only be answered from inside the business, by someone senior enough to ship. So we run both: engineers deployed into accounts to find the workflow worth building, and a factory that builds it on a stack we have already hardened elsewhere. Whatever generalises returns to the catalog, so the next account starts from a platform instead of from zero. Doing only the foundry half is not a smaller version of this. It is half of it.
Proof
Five platforms in the catalog. Four are in first pilot inside partner accounts — LegalOS and Mould in enterprise operations, LedgerVault and the remittance-withholding workbench built on it in financial services — and Keel is in build for real estate. LedgerVault generalised far enough to become its own company and now sits in the portfolio. Partner accounts are described by sector; we do not publish client names.
How we work
The mechanics, in plain language.
- Sprint length
- 2–4 weeks
- Each sprint ends with a deployed environment a paying customer can use, not a slide deck.
- Team shape
- Senior, full-stack
- Most engineers ship across the stack. We don’t hand off between front-end, back-end, and ML teams; one team owns the venture end-to-end.
- Engagement model
- Deployed or cast
- Three new engagements a quarter. Deployed: an engineer works inside your business, and we build what they find on an existing platform. Cast: we build the company ourselves and the studio holds founding equity. Most revenue comes from the first.
- Default deployment
- Vercel · AWS Mumbai · on-prem
- Cloud-native by default; on-prem when the customer’s regulator demands it. Same Kubernetes manifests either way.