Our services

Bearing Activate is a weekly B2B sales intelligence system that monitors a client's full prospect list automatically — no human research required. Every week it pulls signals from RSS feeds, Perplexity search, and website change detection for each prospect, then uses Claude to score each signal for sales relevance, generate personalized outreach copy, and assemble everything into a branded HTML/PDF intelligence report delivered by email every Monday morning.

The engineering challenge was processing 100+ contacts within a practical time window. The system uses async parallel processing across all data sources simultaneously — RSS, Perplexity, and changedetection.io run concurrently per contact rather than sequentially. Result: a full report run completes in under 5 minutes regardless of list size.

Infrastructure: built on GCP with Supabase (Postgres) as the data layer for contact records, signal history, and report state. Self-hosted changedetection.io replaced a $80/month SaaS tool. Claude handles three narrow judgment tasks — signal relevance scoring, outreach copy generation, and report narrative — while all data fetching, scheduling, and delivery runs in pure Python.

Built for an Amazon FBA seller to replace a slow, error-prone manual research workflow. The system ingests product niche data from multiple third-party APIs, runs a multi-stage validation pipeline to score market opportunities, and generates formatted Excel deliverables — cost analysis, opening order model — ready for immediate sourcing decisions. No human touching data between input and output.

The core engineering work was a deterministic, resumable job architecture: each pipeline phase checkpoints its output to SQLite, so a failed or interrupted run picks up exactly where it stopped rather than restarting from scratch. This matters for a research pipeline that calls external APIs with rate limits — a mid-run failure is a cost and time problem without checkpointing.

Claude (Anthropic API) handles three specific judgment tasks where natural language reasoning outperforms pure Python: competitive theme synthesis across product listings, niche scoring against business rules written in plain English, and HTS customs code classification for sourcing. Everything else — data fetching, math, file generation — stays in pure Python. The system writes results back to Google Sheets via a column-ownership model that prevents automation from overwriting cells the client manages manually.