Balloon Sight Intelligence · 2026 – present
A construction-project data pipeline and the product on top of it
Founder and sole engineer

What it does
Nine public data sources feed it: municipal permit APIs over Socrata, ArcGIS REST and OpenDataSoft, the BC Major Projects CSV, the BC Environmental Assessment registry, and federal CanadaBuys tenders. Each arrives in a different shape and is normalised into a single SQLite schema with deduplication and per-source incremental fetch state. Roughly 9,800 lines of Python across 11 agent packages, 131 commits.
Every subscriber gets a weekly digest filtered and ranked to their sector and trades, read from their Stripe profile metadata and generated as Markdown and HTML through the Claude API.
Status: pre-revenue, validating with Alberta and BC contractors. There are no customers to point at.
Operations
Eleven agent packages covering ingestion, contact enrichment, digest generation, outreach and analytics run on systemd timers with OnFailure units that email when something breaks.
Cost was a design constraint from the start: five paid APIs on a $5 VPS, held to $110–150 a month. Per-run budget caps were added after a hung Apify run exhausted the monthly quota and took the enrichment agent offline.
The front end
The customer-facing product is Next.js 16, React 19, TypeScript and Tailwind 4 on Vercel: marketing site, gated free-sample funnel, Stripe Checkout and subscription onboarding, and an internal dashboard for approving outreach before it sends.
Outbound email is CASL-compliant by construction: verified sending domain, physical-address footer, reply-to-unsubscribe, a suppression list checked before every send, and a daily send cap.
Known limitations
Send state is split across two SQLite copies, laptop and VPS, because either can run outreach. The fix is a single authoritative database; Postgres is already installed on the box.
There is no IMAP reply polling, because inbound volume was two emails.
In hindsight the outreach and reply loop should have been built before the intelligence, SEO and community agents. Distribution was the bottleneck throughout, not features.
Python · SQLite · Next.js 16 · React 19 · TypeScript · Tailwind 4 · Stripe · Claude API · Hetzner · systemd