Build plan
Most of the code already exists. What’s left is signing up for four accounts, running two SQL files against Neon, deploying five workers and one dashboard via wrangler. A focused afternoon. Sandwich at lunchtime.
Open four browser tabs, sign up for each, save the keys somewhere safe. None of this needs technical skill.
Add a card. Upgrade to Workers Paid ($5/month). This is the platform that runs everything.
console.neon.tech. Region eu-west-2. Copy the pooled connection string. This is the database.
developer.company-information.service.gov.uk. Free. The crawler uses this for filings, charges, officer changes.
console.anthropic.com. Powers the article classifier. £20 lasts months at typical volume.
This is the technical block. Two SQL files into Neon, five wrangler deploy commands, secrets pasted in. Following DEPLOY.md step by step it takes about two hours including the inevitable typo or two.
Paste the contents of each file into the Neon SQL editor. The schema creates all tables and views; the seed populates 11 sample UK retailers (Debenhams, Carpetright, Cath Kidston, River Island, SecretSales, Next, Currys, JD Sports, Boots, M&S, John Lewis) with realistic-looking signals so the dashboard works on day one.
seed-news-sources-uk-de.sql adds ~100 RSS feeds across UK and Germany. Run the smoke-test script after to disable any that 404. Replaces the 6 starter sources in the main seed.
Companies House crawler. News scraper. LLM classifier. Risk scorer. CSV-ingest. For each: cd workers/<name>, npm install, wrangler secret put DATABASE_URL, wrangler deploy. The CH worker also takes CH_API_KEY, the classifier takes ANTHROPIC_API_KEY.
One-liner each. KV is for the Companies House rate limiter. R2 is where Redbrain finance will drop daily invoice CSVs for the csv-ingest worker.
The last hour. Deploy the dashboard, point a password at it, share the URL. Then spend twenty minutes adjusting the sliders to match Redbrain’s actual risk tolerance.
cd dashboard, npm install, npm run pages:deploy. Set DATABASE_URL and DASHBOARD_PASSWORD as Pages secrets. Redeploy once more so the env vars take effect.
Walk down the fifteen scoring parameters. Half-life. Severity weights. Band thresholds. Action triggers. Defaults are sensible, but Redbrain’s book has a specific shape — nudge until the right merchants surface at the top.
Apply tags — Critical watch, Renewals due, PE-owned. The system starts pulling extra data on the deeper-tagged ones immediately. Confirm they sit at the top of the Finance view.
Add recipients on the Alerts page. The five default rules cover most needs — critical breach, score jump, overdue threshold, watchlist movement, daily digest. Mute the ones you don’t want, route the rest. Done.
A day to deploy. A fortnight for the data to fill in.
The CH worker runs every 6 hours. Within a day or two, every merchant in the book has its complete filing history, charges, officer changes, accounts-overdue flags loaded. Real signals start replacing the seeded ones.
RSS scraping every 30 minutes across ~100 sources. Each article goes through Claude Haiku to extract merchant mentions and distress severity. Press signals stack on top of CH and internal data.
The csv-ingest worker watches an R2 bucket. Drop daily invoice + exposure CSVs in, the worker upserts into Neon overnight. From this point, DSO and exposure data is live not seeded — the most predictive signals in the whole system.
Review the merchants the system flagged versus what finance/credit already thought. Adjust sliders. Reset or add watchlist tags. Calibrate alert thresholds so the team gets pinged for what matters without crying-wolf fatigue.
What was an 8-week plan in the original proposal is now a one-day deploy because the code already exists. The real work is the tuning, not the building.