Problem
Web intelligence tooling is fragmented — scrape here, monitor there, enrich elsewhere — with unclear product boundaries and naming churn that blocks composition.
Approach
Organize the product around five metal-named primitives that each do one job, share auth/data/AI infrastructure, and compose into workflows without big-bang refactors.
Decisions
- Metal names in code (extract, observe, scout, enrich, agent); domain language in the UI — reduces renaming churn.
- Shared authentication, database, and AI stack so primitives stay composable under one billing model.
- Firecrawl-backed extraction and AI SDK agents for interactive research sessions with citations and tool orchestration.
Outcome
Documented platform architecture in the Daedalus README with a working Next.js application surface (dashboard screens captured in UI review). NO DOCUMENTED PUBLIC USAGE METRICS — flag for human if launch numbers should appear.
What I'd do differently
Would publish a short architecture diagram and eval harness next to the primitives table so interviewers can see quality gates without a live demo walkthrough.