A team of AIs, thinking together.
Project brief
Turned opaque AI research into an inspectable agent workflow with live execution state, manual checkpoints, reruns, citations, and exportable reports.
Hiring signal
Senior ownership across product framing, execution scope, and shipped interface quality.
Engineering read
Focus on constraints, architecture bets, external systems, and data-flow decisions.
A guided multi-agent research assistant. A goal goes in, a planner breaks it into discrete tasks, a research router grounds each one in web or reasoning, and a writer/critic loop turns the result into a cited, polished report — running entirely on local LLMs via Ollama.
Most "AI research assistant" demos are a single opaque call to a hosted model: you can't inspect what it's doing mid-run, you can't pause it, and you're paying per token to a third party for something that could run locally.
I broke the pipeline into five inspectable stages — planner, research router, researchers, writer, critic — and built the frontend around a React Flow graph so every stage is visible as it runs, not just the final output. Server-sent events stream state updates live, so the UI reflects the actual execution, not a simulated progress bar.
Project method
The starting requirement was not another chat interface; it was a transparent research workflow where a user could see how an answer was being planned, grounded, rewritten, and critiqued. That pushed the scope toward inspectable stages, manual checkpoints, rerun states, and citations rather than a single prompt-response screen.
Next.js App Router handled the product shell and streaming API routes, React Flow made the agent graph legible, and Material UI provided dense controls quickly. Ollama was chosen deliberately so the core promise could be local inference with no paid model dependency, while Server-Sent Events kept the run graph synchronized without the complexity of a bidirectional socket protocol.
The MVP was defined as one complete goal-to-report loop: planner, research router, editable researchers, synthesizer/writer, critic, and citation-aware output. Features that did not prove that loop, such as persistent multi-user memory or configurable remote model pools, were left for later.
The project moved from a linear orchestrator to a checkpointed pipeline. Once planner output existed, the next critical step was making stale research visible and rerunnable before synthesis, so the UI evolved around graph state, pause/continue controls, and review modals instead of a passive progress indicator.
Conclusion
The result is a pipeline you can watch think: a live graph shows planner → router → researchers → writer → critic, with "needs rerun" badges on stale outputs, pause/continue and auto-run toggles, and zero dependency on a paid API since inference runs fully on Ollama.
Proof list
Open channel
Architecture, AI workflows, design systems, and interfaces that need to hold up after launch.