AI research · Working product
Market Mantis
From chat agent to shared research.
I translated a research agent running in Claude into a shared product with scheduled research, searchable reports, visible sources, and model-cost controls.
01The problem
A client had a capable research agent inside Claude, but it ran only when prompted. Findings stayed in conversation history, other people could not build on the work, and model costs were difficult to inspect.
02My ownership
I owned the product model, information architecture, interaction design, agent workflows, and end-to-end build. The remit was to turn a useful research workflow into a system people could run, review, and revisit.
03Key decisions
Decision 1
Keep evidence beside the answer.
Ask Mantis exposes the reports and sources used in a response. Contextual questions stay beside the brief under review so the reader can inspect the reasoning and keep the final judgment.
Decision 2
Make research persist beyond a conversation.
Scheduled agents publish into a searchable library of reports and signals. The workspace retains the output so people can revisit it and ask follow-up questions across research.
Decision 3
Make operating cost visible.
Usage and spend views show what model-powered jobs cost. This puts operating economics alongside the research experience when considering which agents and models to run.
04Outcomes & next question
- A working platform connects scheduled agent workflows, reports, signals, and evidence-backed chat.
- Research remains searchable beyond the original conversation.
- Model usage and spend can be inspected within the workspace.
Technical details
React · TypeScript · Python (FastAPI) · PostgreSQL (Supabase) · Claude API · Gemini API · Vercel