AI services
AI in production, not in a slide deck
Twelve of our forty-two documented projects put a model, a rules engine or an automated decision into production. This is what they run on, and what we learned shipping them.
Discuss an AI project Read the case studies
- 12AI, rules and automation projects live
- 4model and NLP engines: GPT, Whisper, Daxtra, Stanford NLP
- 5Drools rules engines that analysts edit
- 2016shipping on Java, Python and Node since
What we build
Four kinds of AI work, each with projects behind it
A model call is the easy part. The work is the queueing, the audit trail and the fallback when a call fails.
-
Language models in products
Generation, transcription and assistants, built so a slow or failed call never leaves a blank screen.
-
Document and language understanding
Resumes, inbound messages and forms turned into structured data, with uncertain cases routed to a person.
-
Decision engines
Rules that change weekly, edited by analysts in a spreadsheet, versioned and tested, live without a release.
-
Automated decision workflows
No model at all: rules about who hears from you and when, holding on every run, even the ones that fail halfway.
In production
Four projects, in detail
-
Confirm firstnothing durable changes without the user
nPhase.ai
AI writes the statistical programs behind clinical tables, inside a 21 CFR Part 11 environment.
-
15 min → 30 secfor staff to get an operational answer
Pamper Me
A master agent hands questions to inventory, staff and finance agents in parallel, orchestrated in n8n.
-
6 appsone Whisper voice search across them
Undu AI
Spoken queries become structured searches, in more than one language, across a supplier marketplace.
-
v1 → v4rule sets changed by analysts, no release
Aclaro
Vehicle recommendations and discounts by VIN, from a Drools engine configured in Excel.
How an AI engagement runs
Start with the decision, not the model
- 01
Find the decision
No decision that changes, no project. Better to learn that in a week.
- 02
Prove it
A narrow proof of concept on your real data, not a demo set.
- 03
Price the mistakes
How often it is wrong, and what each kind of wrong costs.
- 04
Build the plumbing
Queues, retries, audit trails, cost controls and a human override.
What we push back on
Four AI projects we will talk you out of
- Replacing a working, auditable rules engine with a language model.
- A model making a decision with regulatory consequence, and no human in the loop.
- Any build where nobody agreed up front what accuracy counts as success.
- Training a custom model when a hosted one with good prompting would do. That is most of the time.
And on rules engines: a spreadsheet with four hundred rows is a codebase without tests, so we build rule versioning and testing in from the first day.
Tell us what you are building
A system to build, replace or rescue. We reply within one working day with who would work on it.