Dallas Crilley
Work
Applied AI enrichment Shipped

EnrichCRM

CSV in, enriched contacts out, at a fraction of the cost.

A contact-enrichment pipeline that turns a CSV of business emails into enriched records, built to do it at a fraction of managed-service cost.

language
TypeScript (Node CLI + Next.js 15)
CLI stack
Brave Search, Trafilatura, Gemini Flash
UI stack
OpenAI agents, Firecrawl, SSE streaming
storage
SQLite, per-contact cost tracking
ownership
Sole engineer — built end to end
history
built Jun 2025 to Apr 2026
repo
Private

The problem

Researching each contact’s company, industry, and tech stack by hand is slow, and managed enrichment services charge by the record. EnrichCRM treats that cost as an engineering problem: a local scraper and a cheaper model in place of the paid stack.

How it's built

  1. Cost treated as an engineering problem

    The CLI replaces a paid scraping layer with Trafilatura (local, no per-page network round-trip) and swaps GPT-4o for Gemini Flash. Measured spend lands near $0.003 per contact in the --stats output. The broader "4.6x faster, 77% cheaper" comparison is self-reported, with no benchmark harness in the repo.

  2. Resume-safe batch processing

    Status is written to SQLite before and after every contact, so an interrupted run restarts at the exact failure point with --resume, and --retry-errors reprocesses only the failures rather than the whole list.

  3. Multi-agent field routing on the UI path

    The web surface categorizes each requested field into one of six domains and dispatches to a purpose-built agent, with cross-source confidence scoring and structured output enforced through Zod schemas.

By the numbers

$0.003 per contact on the CLI path (traceable in --stats)
4.6x faster than the managed path self-reported
6 specialized enrichment agents on the UI path

self-reported marks figures stated in docs or commit history that the source brief couldn't reproduce from the repository alone. Everything else traces back to code.

  • 4.6x: Basis: README comparison of developer-observed runs: Firecrawl + OpenAI at 47.6s and $0.013/contact versus Trafilatura + Gemini at 10.4s and about $0.003/contact. The repo tracks CLI cost via --stats, but does not include a benchmark harness for the timing multiple.

Where this honestly stands today

Shipped. Two enrichment surfaces (a Brave + Gemini CLI and an OpenAI + Firecrawl web UI) share types but not providers, and the README is candid about the seam. There is no spreadsheet UI and no Salesforce integration.

Want the parts that aren't in a public repo? Happy to walk you through the architecture and the decisions on a call.