Slingshot — AI Code Copilot
An enterprise AI copilot taken from proof of concept through MVP to delivered product — grounded in internal code and documentation, with retrieval respecting existing access boundaries.
- Scope
- Internal code + docs
- Stage
- POC → MVP → product
The problem
Engineers were losing time re-discovering things the organisation already knew — how a service was wired, why a config existed, where a pattern was already implemented. A general-purpose assistant could not help, because the answers lived in private code and internal documentation.
The approach
I worked on the copilot across its full lifecycle — first as Sage, later renamed Slingshot — delivering the proof of concept, the MVP, and the successive updates through to the shipped product. That meant owning the RAG architecture, the data pipelines, and the model tuning:
- Secure retrieval across internal code and docs, with access controls enforced at retrieval time — a user’s results never include sources they could not otherwise open.
- Response grounding, so answers cite and depend on retrieved internal context rather than the model’s own priors.
- Data pipelines keeping the index current as code and documentation moved.
- Model tuning to fit the assistant to internal terminology and conventions.
- Grafana instrumentation so retrieval and response quality were observable in production, not inferred.
The outcome
The copilot became a working entry point into internal knowledge — engineers could ask about infrastructure and get answers anchored to the organisation’s own sources. Getting it past the prototype stage came down to solving the two things that usually stop these systems: access control and grounding.