LankaBangla Securities: A Governed, Air-Gapped LLM Layer for Capital Markets

Ternary is designing an air-gapped, governed LLM application layer that lets brokerage advisors query market and portfolio data in natural language — without exposing production systems to free-form AI.

The story

How we approached the work, what we built, and why it matters.

The challenge

Relationship managers, brokers, and dealers need fast, conversational access to market data and client portfolios — but giving a language model direct, unconstrained access to production financial systems is not acceptable in a regulated capital-markets environment.

Our approach

Ternary is designing a governed application layer around an air-gapped model — provider-neutral by design, so it can run on a managed service or an open-weight model on dedicated infrastructure. The model never touches raw systems; it works through a curated catalog of approved questions behind a strict data boundary.

What it includes

The design pairs a deterministic validation layer with a secondary "LLM-as-judge" check, role-scoped access so each advisor only reaches data they are permitted to see, and full audit logging. A web workspace gives advisors a chat experience, a discovery view, and saved insights, with safety benchmarked against a purpose-built test set.

Where it stands

The engagement is in its design and architecture phase, with the application layer and governance model defined ahead of build.

Related case studies.

All case studies

Have a similar problem worth solving?

Tell us where you’re headed. We’ll bring the engineering discipline to get you there.

Get in touch