Visca

AI / ML teams

Ship from notebook to production on one in-perimeter ecosystem — open-weight models and all the surround, no stitching.

You build the agents. What slows you down isn't the model — it's everything around it, each a different tool: identity, secrets, eval, deploy, observability — and in a regulated shop, all of it has to run self-hosted with open-weight models in-perimeter so no prompt or output ever leaves. Visca gives you one ecosystem for the whole loop instead of a stack you stitch and re-stitch, so the thing that worked in the notebook is the thing that ships, inside your own walls.

Why the data can't leave

Hosted AI is out. A stitched stack is the only thing left — and it breaks here.

The operational surround is a second project

A working agent in a notebook is 20% of the work. The other 80% — identity, secrets, eval, deploy, monitoring — is a different stack you didn't sign up to build.

Evals that don't match production

You test against a harness that drifts from how the agent actually runs. Regressions surface in production, not in CI.

Reproducibility you lose by Friday

The model version moved, the prompt changed, a tool updated. The agent that aced eval Monday can't be reconstructed by the time it regresses.

One ecosystem, not a stitched stack

One self-hosted ecosystem, applied to ai / ml teams.

Praxis

One loop: author, eval, deploy, observe

Bring up the full substrate locally in under a minute. The same SDKs and primitives run from laptop to production — no rebuild at the deploy boundary.

Seal

Reproducible to the byte

Every agent is a content-addressed, signed bundle pinning model, prompt, tools, and dependencies. Monday's eval-winning agent is one hash away, forever.

Lattice Runtime

Production-shaped runtime

Token budgets, model fallback, durable execution — the operational realities, handled by the runtime instead of your application code.

Chronicle

Replay with token-level granularity

Every run recorded and replayable. Debug a regression against the exact context that produced it.

What you get

Outcomes.

In practice

From notebook to production in one surface

You prototype an agent locally with `visca init` — the full substrate, in under a minute, no cloud signup. You eval it against a declarative spec, package it as a signed Seal Bundle, and deploy to Lattice Runtime with one command. The agent that passed eval is, byte-for-byte, the agent in production — and every run is replayable in Chronicle.

Other roles

The whole stack. Self-hosted. One ecosystem.

The entire agent stack, inside your own walls.

Models, identity, tools, voice, payments, runtime, and audit — as one integrated ecosystem, self-hosted, sovereign, air-gapped. Nothing stitched from vendors. Nothing leaves your perimeter. Open at the core. No license rug-pulls, ever.