Gleaned: a live offers platform, built with AI in days, not months

The idea

Organic farms selling direct run genuinely good offers — but they're scattered across dozens of farm-shop newsletters nobody reads. Aggregators solve that by taking the customer relationship (and a cut). Gleaned takes the opposite side: "Gleaned is on the side of the people who grow your food." It watches the online shops of certified organic producers daily and sends subscribers one clean digest — every offer linking straight to the farm's own shop, so the sale, the customer and the margin stay with the farm.

Buy Local. Buy Direct. Buy Organic.

The product

Subscribers set it up once, then read one email a day. Behind that simplicity:

• User accounts — email/password and Google sign-in, with a live dashboard for digging deeper than the digest
• A morning digest that only arrives when something genuinely new appeared — no empty emails, ever
• Price honesty — every day the pipeline compares prices against history, normalises per-kg and per-bottle costs, factors in delivery thresholds, and keeps only genuine discounts, multibuys and price drops

The engineering

Robust ingestion, not fragile scraping. Wherever possible Gleaned reads structured store APIs — Shopify products.json and the WooCommerce Store API — rather than parsing fragile HTML. Retailers that block automated access entirely (Waitrose and Ocado block at connection or datacentre-IP level) are covered a smarter way: their own offer newsletters, read from a Gmail inbox via the Gmail API.

A daily pipeline. A GitHub Actions run at 07:00 UTC checks every shop, searches the inbox, and diffs everything against the previously-seen offer set stored in Netlify Blobs. Only genuinely new offers survive the diff — which is why subscribers trust the email.

The sanity gate. The part most automation skips: before anything is persisted or emailed, every run passes a three-level correctness gate —

• Per offer: implausible per-kg prices (weight mis-parses), deposit wording and too-good-to-be-true ">75% discounts" are quarantined out of the digest
• Per farm: if a farm's median price drifts more than 40% against its last-known-good result, the pipeline serves the last good catalogue and refuses to save the anomalous scrape as the new baseline
• Per run: a sudden flood of "new" offers — the signature of a diff bug or a site-wide re-list — holds the entire run: nothing committed, nothing emailed, alert sent. A later clean run re-detects everything safely

AI adjudication, used correctly. Quarantined items get a second opinion from a small Claude model: false positives are released (with corrected weights where inferable), and every verdict is cached so each item is judged exactly once. The AI unblocks the automation — it never overrides the gate.

The same pattern runs a US pipeline against American organic retailers, sharing the sanity-gate architecture.

How it was built

Honestly: with AI, and fast. I directed; AI coding agents (Claude) built alongside me. The product decisions, the ethics of the model, the correctness rules in the sanity gate, the choice of what never to automate — those are mine. The implementation — scrapers, pipeline, accounts, dashboard, the lot — was pair-built with AI in days, not months, at a cost that would once have been a single agency workshop.

That's not a footnote; it's the demonstration. When I sell a prototype sprint — "a working internal tool, deployed in days" — Gleaned is what the method produces when it's pointed at a real market. AI-assisted building only works when someone with domain judgement is steering; this whole platform is what that looks like in practice.

The prospecting engine behind it

Gleaned's sister system answers a different question: who else should we be talking to? Starting from the Defra organic register — the public record of every certified organic operation in the UK — an enrichment pipeline cross-checks each producer through three independent search APIs (Google Custom Search, Serper and Bing), re-verifying every match before it's accepted so directory false-positives never pollute the data. The result is a segmented prospect database — producers with shops to monitor for Gleaned, brands for Organic Guardians consultancy, and suppliers for Skär Organics — built and maintained by one person, because the verification is automated and only the judgement isn't.

Why it's on this site

Three reasons. It's shipped product — not a demo: live at gleanedorganics.com, running daily, with real subscribers and 23 producer relationships. It's the working proof behind my AI marketing automation offer — the monitor-diff-digest-gate pattern transfers to any market: competitor pricing, industry job boards, supply markets. And it's part of the organic-sector ecosystem I've built as a certified organic producer myself — alongside the producer database built from the Defra organic register and Organic Guardians, the consultancy it all supports.