Overview

CustomGPT is a crypto market-intelligence assistant: a chat application whose answers are assembled, on demand, from live market data rather than from a model's memory. A trader asks "what's the funding rate on BTC futures?" or "show me whale movements and the long/short ratio" in plain language, and the assistant decides which of sixty-seven declared tool functions to call, fetches the real numbers, runs the indicator maths and answers with charts and tables.

Underneath the chat sits an integration layer spanning five market-data providers — spot and futures pricing, derivatives positioning, market-cap reference data, on-chain metrics and social sentiment — declared as one function schema an LLM can reason over. On top of it sits a Streamlit workspace with sign-in, per-user chat history, file upload and export, backed by a small relational store.

I joined as a contract engineer for the final push. The product existed, but the work of three people sat on three divergent branches that no longer merged, and a large part of the tool surface either failed outright or returned data the assistant could not use. My brief was narrow and concrete: make it one codebase, make every function work, and prove it.

Problem

An LLM tool surface fails differently from an ordinary API, and worse.

When a normal endpoint breaks, a caller gets an error and handles it. When a tool function breaks, the assistant improvises around it — it answers anyway, in confident prose, with whatever it can reach. A partly-broken market-data surface therefore does not produce visible failures; it produces plausible, wrong financial answers. That made "most of the endpoints work" an unacceptable finishing state.

The integration surface was also unusually wide for its age: five vendors, each with its own auth, rate limits, pagination, symbol conventions and failure vocabulary. The same asset is spelled differently by each provider, and several of them answer a bad request with a 200 and an empty body rather than an error — so "it returned something" was not evidence of anything.

And the code was in three places at once. Three contributors had worked in parallel on branches that had drifted past the point where the tooling could reconcile them; the histories overlapped, the same functions had been edited in incompatible ways on each side, and there was no test suite to tell a good merge from a bad one. Merging blind risked silently reverting someone's fix — which, on this surface, would look exactly like nothing at all.