Live
Ingest Lens
Data never arrives clean. Build for that.
Operational information is usually fragmented across documents, forwarded email, forms, notes and specialist tools. The value is not extraction for its own sake, or another interface: it is bringing that information into a useful hub, connected to the people and work it affects, so routine transfer happens automatically and human eyes are reserved for exceptions, judgement and action. Ingest Lens takes one deliberately awkward document through that intake edge and keeps every transformation inspectable.
This is the visible intake edge of systems such as Impact Desk and Production Desk — not a standalone destination for the data.
The opportunity
Designed for Production companies, nonprofits and small organisations whose real inputs are scans, forwarded email, business cards and handwriting
Systems that assume clean data push the cleaning onto a person. The cost is not filing — it is the hour that belonged to a client, and the follow-up that never happened because nobody had time to type it in.
Working now
- ✓Five inspectable stages over a real scanned document: identify, extract, repair, structure, review
- ✓A fallback ladder that records why each rejected method failed, not only which one worked
- ✓Every field traceable to the pixels it came from, or explicitly marked as untraceable
- ✓Each API stage run on three models, with tokens, cost and per-field agreement shown
- ✓Parses into the real schema of a working CRM rather than an invented one
Next on the roadmap
- 01A vendor quote arriving in response to an automated RFQ, where several people hold one piece each
- 02Spoken capture: one utterance containing a deadline, a reminder and an idea, proposed as separate items
- 03Reading a page of your own notes into project points
- 04The same review surface generalised to approve items bound for different destinations
Project in brief
“You are at a conference and you meet someone worth following up. You take a card, or a resume. And until now, everything after that moment was on you: find the time, sit down, retype it into something that can chase it, reconcile it against your notes and whatever else is competing for the week.
That is the part software usually leaves alone. A database with a good web interface is the core, and the core is the easy half. The hard parts are the edges — data arriving in whatever shape it arrived in, the right thing surfacing at the moment it is needed, and a person spending their attention on a decision rather than on being the manual interface between two tools that each work perfectly.
So this demo takes one deliberately awkward document — a resume that exists only as a photograph of a page, two columns, four date formats, a table of credential numbers — and shows every stage of what happens to it. Where the naive method returned nothing and why. What the OCR actually read, alongside what was corrected. What each model cost, and where a model a seventh of the price was just as good, and where it was not.
Nothing in it is marked as verified. Every field either points at the pixels it came from, says it was corrected and shows what was underneath, or admits it cannot be traced at all. That last category is the point: a system that hides its uncertainty is asking you to trust it, and a review step only means something if disagreeing with it is easy.
The document here is one example, and so is where it lands. It could arrive from a watched mailbox or an upload page; it could be speech, or your own notes. Those are details of a deployment. What does not change is the assumption underneath: the data will not come in clean, and the person should be reviewing it rather than assembling it.”