For regulated AI teams
We build your firm its own model, tuned to your book, cheap enough to own and run, with the model-risk evidence your committee and counterparties require. We pick the right model for your task and train to it. Our method re-targets across models as the field moves, so you are never stuck on one.
A short technical briefing on Tensyl Foundry. No commitment.
Anthropic partner model builds are provider-flexible across Anthropic, Microsoft, and Fireworks
Built for SR 11-7 · NAIC · EU AI Act · FDA PCCP
The problem
Frontier API bills scale with every token and never stop. You never own the asset, and you cannot fully control or inspect it.
Your evals are your IP
Rent, and your proprietary logic for what counts as a good answer lives inside someone else's closed model.
An off-the-shelf model is not tuned to your book, so it underperforms on your lines and your language.
Your model-risk committee will not clear it without provenance and evidence a research boutique will not produce.
How it works
Your data, under your controls. We define the use case and the acceptance bar together.
We choose the best-fit model for your task and budget, from any provider, then train to it.
We find the signals in your book that a general-purpose model misses.
The evaluator that encodes your standard, plus the evidence pack, deployment-ready and handed to you.
Your data and method stay portable across models. The governance evidence is forged in as first-line documentation, not bolted on after.
Build vs. buy
Standing up a model on Bedrock or Azure OpenAI with your own governance layer is a real option, and the right one until you need owned weights, regulator-grade evidence, and a cost you control. That is the line Foundry is built for.
| Build it yourself | With Foundry | |
|---|---|---|
| The asset | You rent API access. No weights you own or can inspect. | You own the weights, running in your own cloud. |
| Fit to your book | A generic model steered with prompts and RAG. | Tuned to your data, your lines, and your language. |
| Model-risk evidence | You build lineage, evals, and monitoring yourself. | Evidence pack forged in as first-line documentation. |
| Committee sign-off | A bespoke package assembled for each review. | One accepted format, reusable across every version. |
| Cost curve | Per-token bills that scale with usage, forever. | A small model you own, cheap to run at volume. |
| Provider risk | Tied to one vendor's roadmap, pricing, and deprecations. | Portable across models as the field moves. |
| Team & time | Ongoing load on your ML, risk, and platform teams. | About a quarter, built with your team, handed to you. |
Table 2. Where an owned model earns its keep. Source: Tensyl, Sep 2026.
Who builds it
Co-Founder & CEO
Product head, Uber Infrastructure and Director at Twitter Growth; product leader for Bing Ads API & platform. VP Product leadership at Instawork (Series D, Benchmark-backed). Rebuilt an operating model around AI and decoupled revenue from headcount.
LinkedIn ↗
Co-Founder & CTO
CTO at Instawork (Series D, Benchmark-backed), from Series A through Series D. Two-time founder with multiple exits; an early pioneer in social and mobile gaming, later CTO of a multibillion-dollar global gaming business. Turned a ten-million-professional labor marketplace into a training ground for real-world robotics.
LinkedIn ↗The evidence pack
One accepted format, reusable across counterparties and across every version of your model.
See it for real
We’ll send a redacted example: model lineage, an independent-eval summary, and a change-control page. You will see the exact format your committee would sign off on, before you talk to us.
Redacted and illustrative. Sent by a person, not an automated download.
Continuity by design
You keep the weights, the evals that encode how your firm judges a good answer, the data lineage, and the full evidence pack. Those evals are proprietary logic you would never hand to a closed API, and here they stay yours. If Tensyl ever goes away, your model keeps running in your own cloud and your committee keeps its documentation. No lock-in, and no custodian but you.
Request a briefing
Request a short technical briefing. We’ll walk through Foundry on one of your use cases and what “accepted” would mean for your committee.