AI research pipeline and CRM
- Segment
- Research tooling
- Role
- Sole developer at Skillion
- Period
- 09 – 10.2026
The team needed to know which of more than 3,200 firms were worth contacting. I built a research pipeline where AI agents look each firm up on the web, rate how well it fits, and write full profiles for the best matches, with every answer saved alongside its sources.
I then set up a self-hosted CRM for the outreach and connected it to Claude, so the team can log emails and follow-ups just by talking to it.
Key features
- AI research at scale. Thousands of firms researched by AI agents running in parallel with live web search. Every prompt, source and answer is saved, so each result can be traced.
- Evidence-tested prompts. Each new prompt version was replayed against saved sources and known-good examples before use. Versions that broke a correct answer were rejected.
- Multi-model and cost control. Batches ran across Claude, DeepSeek and GPT models with automatic switching under quota limits, and a cost guard stops any run at a set budget.
- Quality checks. A second AI pass re-checked the top-rated firms and caught ones that had closed or merged. A hand audit of a random sample found 86% of fields fully correct and under 2% wrong.
- Full profiles. For the best matches, the pipeline crawled each firm's website and wrote a profile. Over 99% of quoted facts were traced back to the saved pages.
- Team spreadsheet sync. Results sync to the team's Google Sheet without overwriting the team's own edits.
- Self-hosted CRM. Twenty CRM on company infrastructure, with custom fields, separate pipelines, and all existing research and contacts imported.
- Automated follow-ups. Workflows update a firm's stage when outreach is logged, create follow-up tasks, flag overdue firms and handle handovers.
- Claude as the CRM interface. Team members connect Claude to the CRM with their own login. A custom skill lets them log messages, see what's due and hand off firms in plain English.
Tech stack
- AI
- Claude (Haiku, Sonnet), DeepSeek and GPT models through OpenCode, Claude via MCP
- Backend
- Python and Node.js scripts
- Data
- PostgreSQL, Redis, JSON and CSV archives, Google Sheets
- Integrations
- Exa search API, Firecrawl, Twenty CRM (GraphQL, REST, MCP)
- Infra
- Dokploy, Docker Compose, SOPS