AI that makes it past the demo.
Most AI projects die between the demo and the deploy. The prototype answers questions well in a meeting, then falls apart on real data: retrieval pulls the wrong documents, the model invents figures, the API bill triples in a week. The gap is rarely the model. It is everything around it.
We build the parts around the model. Search over your own data, with the retrieval step measured rather than assumed. Document pipelines that pull structured fields out of invoices, contracts, and forms, and flag the ones they are not sure about. Forecasting from your sales or sensor history, with the expected error stated up front. Everything runs on OpenAI or Anthropic models behind hard cost ceilings, so a bad loop cannot burn a month's budget overnight.
You get one fixed price in writing within 48 hours of a scoping call, a live URL by the end of week one, and a written update every day. MVPs go live in 2–4 weeks; custom business systems take 3–8. The repository sits in your GitHub organisation from day one.
Startups adding an AI feature to a product, or building the product around one: semantic search, a support assistant grounded in your docs, extraction from user uploads. You do not need a data science team to start. You need one working feature with a cost ceiling.
Established businesses sitting on years of documents, tickets, or transactions and getting no answers out of them. Typical first projects: pulling line items from supplier invoices, forecasting stock by branch, making a decade of PDFs searchable by meaning instead of filename.
Also teams that already tried. If an internal experiment or an agency prototype stalled, we can take it over, measure where it actually fails, and either fix it or tell you plainly it is not worth pursuing.
Scoped in writing. Delivered in your accounts.
Search over your data
Semantic and keyword search combined, with retrieval quality measured against a test set before launch.
Document extraction pipelines
Structured fields pulled from invoices, contracts, and forms, with a confidence score on every extracted field.
Forecasting from your history
Demand, revenue, or sensor forecasts built on your own records. We tell you the expected error before you rely on it.
Hard cost ceilings
Every OpenAI or Anthropic call runs behind a spend cap, so a usage spike cannot run up the bill.
Evaluation before launch
A test set built from your real cases, run on every change, so quality regressions surface before users see them.
Human review queues
Anything the model is unsure about lands in a queue for a person to clear before it reaches your records.
Background job plumbing
Batch work like overnight extraction runs on Node with Inngest: retries, dead-letter queues, no dropped documents.
Your repo from day one
Code lives in your GitHub organisation from day one, with full IP assignment.
Scoping call and NDA
A mutual NDA is signed before you share anything, same day, as standard. Within 48 hours you get one fixed price in writing, never hourly, and the number does not change after kickoff.
First working slice
We build the narrowest end-to-end version first: your real data in, real answers out, on a live URL you can click by the end of week one.
Measure and harden
We build the evaluation set, tune retrieval or extraction against it, and wire in cost ceilings and review queues. You get a written update every day, and the URL stays current.
Launch and 30-day cover
It goes live in your infrastructure. For 30 days after launch, anything we built that breaks is fixed at no cost. After that, optional monthly care you can stop any time.