Project · Product
A collaborative AI workspace for conversations, models, and workflows — with lockdown mode so data processing stays inside a region you choose.
Most AI chat UIs are thin clients over a single vendor API. Chisel is built as a product surface for teams that care where inference runs: pick models, keep conversations and workflows in one place, and constrain processing geography when compliance requires it.
It sits next to the Diabase open-weight models — the same person building both the application layer and the models that can ground it. That end-to-end ownership is the point of this write-up.
Scope
Multi-user oriented surface for managing conversations, model choice, and workflows in one product — not a one-off demo notebook.
Designed to work with models you choose, including Diabase open weights and external providers, rather than locking the product to a single closed API.
Restrict data handling and LLM usage to a configured country or region. When enabled, processing stays inside that boundary for the workspace — including all members.
Content belongs to the user. Diabase does not use workspace content to train models. Retention and deletion follow the published privacy policy.
Architecture
At a high level, Chisel is a full-stack AI application: a web client, an application/API layer, persistence for accounts and conversations, and adapters to inference backends (hosted and/or self-hostable).
Lockdown mode is a first-class product constraint, not a marketing footnote — provider and region selection are part of the control plane, so compliance choices show up in what the system is allowed to call.
Decisions
Lockdown mode encodes jurisdiction in runtime behaviour. The alternative — “we’re GDPR compliant” with no enforceable region switch — doesn’t help a team that must keep processing in-country.
The product is oriented around selectable models and open weights (including Europa and the retrieval stack), so the app remains useful if a provider changes terms or a team wants on-prem inference.
Priority went to a usable multi-user product surface and policy controls first. Retrieval models are published as research previews in parallel — see the retrieval case study — so RAG can deepen without blocking the application.
To harden toward a commercial product: reliability SLOs, richer admin/audit, production RAG with hybrid search, formal security review, and clearer tenancy/billing. The independent Diabase work is the lab that feeds that roadmap — not a claim that the product is finished.
For hiring managers: Chisel is evidence of product engineering across the AI stack — UI, backend concerns, model integration, and security/compliance product decisions — not only cluster operations or a single fine-tune script.
Try the live app at chisel.diabase.ai. If the host is slow or unavailable, this page is intentional: a recruiter should still understand the system without creating an account.
