The Model Is the Easy Part: Lessons from an Air-Gapped LLM in Capital Markets
LankaBangla Securities had rich client and market data — and no way to reach it without a technical intermediary. Getting AI into that environment was never a modeling problem. It was an architecture problem.
Two gaps, one constraint
The firm faced a dual gap. Dealer-brokers spent too much time gathering and synthesizing dispersed information before they could act. Retail marketing, meanwhile, risked irrelevance whenever messaging wasn't tied to account state and trading behavior. And over both sat a hard constraint: in a regulated capital-markets institution, free-form AI access to production data — or sending sensitive data to public AI endpoints — was never an option.
The boundary comes first
The answer started at the boundary, not the model. An open-source LLM was deployed in an air-gapped environment, so inference — and the sensitive data feeding it — stays inside the institution's own walls. With boundary control solved, the real work began: an application layer that mediates every interaction with the model.
A governed layer does the heavy lifting
That layer is the control plane. Natural-language requests resolve to pre-approved, read-only query templates; the model never touches production data directly; role-based visibility decides who sees what; and every interaction is logged. The model provides capability. The layer provides trust.
Ship something concrete, early
Two functions went first: an assistant that lets dealer-brokers retrieve and synthesize information inside their decision windows, and marketing automation for retail traders that adapts messaging to trading behavior and account context. Both run on the same platform substrate — which means the next function reuses the governance, orchestration, and integration patterns instead of rebuilding them.
What we took away
In regulated environments, AI adoption succeeds when architecture and workflow design are treated as one problem. Air-gapping addresses boundary control, but the business value comes from the layer that converts model capability into role-specific, auditable execution.
The second lesson is sequencing. Shipping two concrete functions early, on a reusable core, creates momentum while preserving long-term scalability. The institution got practical AI without regulatory exposure — and a foundation ready for whatever function comes next.