Loading…
I'm Folarin. I build AI systems that survive contact with production.
Nine years building software, the last two almost entirely on getting agent systems to behave once real users depend on them. That work is mostly routing, retrieval, evaluation and oversight, and it looks a lot less like prompting than people expect.
A demo proves a model can do something once. Production is the argument that it will keep doing it, at a price you can defend, in a way someone is willing to sign off on.
- Senior Product Engineer on the AI platform at Cuvama, since 2024. I own the frontend platform and the product's AI surface end to end.
- A tiered multi-agent backend in LangGraph that routes each task by complexity, cost and token budget across OpenAI, Anthropic and Bedrock. Cut report generation time by more than half and LLM spend by 40%.
- Version-aware retrieval over enterprise knowledge: chunking strategy, embedding choice, and Agentic RAG on MongoDB Atlas Vector Search. Typed tool calls replaced bulk context injection, cutting per-request tokens 35% and measurably reducing hallucinations.
- LangSmith evaluation pipelines on curated datasets, wired into GitHub Actions so an eval regression blocks a merge the same way a failing test does, with alerting on content-quality benchmarks in production.
- Plan-approval workflows for consequential changes: agents propose multi-step edits, people review and revise section by section with versioning and regeneration, and nothing commits unreviewed.
- Streaming chat, tool-call and agent-activity visualisation, and progressive rendering over WebSockets so a pipeline that runs for minutes stays legible while it works.
- Agentling, a Python framework for reliable agents: live streaming of actions and tool use, automatic recovery from failed steps, and persistent memory so runs can be paused, resumed and replayed.
- Five years at lead level, including frontend teams of seven and five, building real-time financial interfaces where every action had consequences: crypto portfolio analytics and live equities trading.
- MSc Software Engineering, Distinction, University of Westminster. Fellow on Nebius Academy's AI Performance Engineering programme, covering LLM architecture, inference optimisation, post-training and MLOps. I mentor engineers at Westminster and on ADPList.