Fintech Firm Overcomes Engineer Departure with Direct Client Engagement in APAC

A fintech intelligence company in Singapore was building an AI-powered financial analysis product for enterprise clients: banks, investment funds, and insurance groups. The product needed to answer complex research questions: governance assessments, accounting anomaly detection, multi-source company comparisons. Each question required searching the web, querying proprietary financial databases, cross-referencing historical studies, and generating charts to support the findings.

The technical challenge was real. A single general-purpose AI model could not handle this level of multi-step reasoning reliably. But the deeper challenge was commercial. Enterprise finance clients have one question before any other: what happens to our data? Who can see it? How is access controlled? Without a credible answer to that question, the product could not be sold.

The company needed engineers who could build both: the intelligence layer and the security layer, and who understood why the second one mattered as much as the first.

Talex deployed two engineers into the client's team: one AI specialist and one backend developer, working alongside the client's own data analyst as the core technical function. Part of a broader four-person engagement with the same client across multiple products, these two focused entirely on this financial analysis system for the duration of the engagement.

They built a multi-agent architecture at the core: a Supervisor agent that receives a research prompt, breaks it into sub-tasks, and delegates to four specialists: a search agent for external information, a query agent for financial databases and research studies, a chart agent for visual outputs, and a consolidator agent that synthesises everything into a single coherent answer. Each agent does one thing well. The Supervisor ensures the parts become a whole.

On top of the intelligence layer, they built an MCP (Model Context Protocol) server: a standalone, secure wrapper that exposes the system's capabilities to external enterprise platforms without touching core logic. Every client connection is isolated: Firestore-based API key authentication, IP whitelisting, Redis-backed rate limiting, circuit breakers, and per-tenant token usage attribution. Enterprise clients, including platforms running ChatGPT Enterprise via GPT Actions, can plug in directly without the vendor losing control of how the system is used.

The engagement ran for two and a half years. When one of the engineers decided to leave, the client reached out directly to ask them to stay. That is the kind of relationship that does not happen with contractors. It happens with people who are genuinely invested in what they are building.

The system handles complex, multi-step financial research tasks autonomously: from web search through database query to chart generation to synthesis, without manual handoffs at each step. The quality reflects genuine domain depth, not surface-level summarisation.

The MCP server changes what the product can do commercially. The system is no longer a single web application. It is a capability that any enterprise platform can access securely, with full usage tracking and without the vendor losing control. New clients and new platform integrations do not require rebuilding the core system.

For a product category where enterprise trust is the primary barrier to adoption, the security architecture is not a feature. It is the product.