Consulting since 2009
Most AI projects are data projects in a better suit.
We help teams find where AI genuinely helps, build the systems that deliver it, and keep the data underneath in good order. We started with the last part, in 2009.
- day 01gjh.discoveryshadow 3 workflows · interview 6 practitioners✓
- day 04gjh.dataprofile source systems · row counts, freshness, null rates✓
- day 09gjh.analysisscore 11 candidates → 2 build · 3 defer · 6 decline✓
- day 14gjh.deliveryrecommendation memo + reference architecture✓
- day 15clientdecide: proceed, defer, or stop — no obligation
What we do
Four practices, one team
Senior people do the work. You will not be handed to a team you have not met.
- 01
Advisory
Where AI is worth the effort, and where it isn't.
Read more → - 02
Building
Assistants, automations, and internal tools that hold up in production.
Read more → - 03
Data foundations
Warehouses, pipelines, and models — the part everyone skips.
Read more → - 04
Staying with it
Monitoring, tuning, and support after launch.
Read more →
How we work
Three commitments we put in writing
- Start small and paid
- A short assessment of one workflow, so you can judge the work before committing to more. Most engagements begin under a month.
- Senior people do the work
- The person in the first meeting is the person writing the code. No handoff to a bench you have not met.
- You own everything
- Code, infrastructure, and documentation, in your accounts. No dependency on us by design.
Industries
Where the constraints do the talking
Across financial services, manufacturing, healthcare, and more — the constraints change what we build.
- Financial services
Banking, payments, and the parts an examiner reads.
Most of what makes AI hard in a bank is not the model. It is entitlements, lineage, and being able to reproduce an answer six months later.
Read more → - Energy and utilities
Meter data, forecasting, and reporting a public board will read.
Interval data arrives late, arrives wrong, and gets restated. Everything downstream has to survive that without quietly changing last month's numbers.
Read more → - Manufacturing
Shop-floor data, quoting, and planning that has to survive the shift.
The skepticism is earned. A demo that works on clean data needs to survive the reality of a floor where the ERP, the MES, and the spreadsheet that actually runs the line disagree.
Read more → - Healthcare
Clinical-adjacent systems where accuracy outranks speed, and privacy is base load.
Patient data is the uncomfortable truth of everything we build here. It sits under the strictest regulations, hides in systems nobody owns to name, and refuses clean structure.
Read more → - Field services
Dispatch, appointments, and the uncertain wait time.
The field lives in the margin the demo forgets. Travel, appointments, and crew availability all move hour to hour, and no model can change the fact that yesterday's route sometimes lies.
Read more → - Construction and engineering
Estimates, schedules, and submittals — the jobsite runs on lagged documents.
Construction runs on the version of the drawing that's current, and the version isn't always the one in the field notebook. The schedules read like fiction to the person on site.
Read more → - Automotive
Dealer data, warranty, and the cost of a part arriving late.
Dealer data lives behind a DMS the dealership does not own, filed by model and part, with the ordering clock the network feels long before the OEM does.
Read more → - Retail
Items, inventory, and the lane where the forecast meets the footfall.
Home goods, grocery, and multi-unit retail are defined by the asset they can't see: the lane, the promo, and the spread that tightens when a promotion goes sideways.
Read more → - Ecommerce
Catalog, conversion, the customers in the browser.
Ecommerce is where 'API-first' meets the client that's a browser tab. The data is everywhere and it is nearly all marketing-shaped until you ask how much it produced.
Read more → - Real estate and rental
The listing, the lease, and the property that no two systems describe the same way.
Real estate data is the same asset described three ways, by a listing platform, a property manager, and a lease — and none of them share a primary key.
Read more → - Education
Enrollment, retention, and the systems that treat a student as several different records.
A student is a fan fold of records — admission, registrar, financial aid, LMS — that do not agree on who they are looking at.
Read more → - Travel
Reservations, demand, and a schedule that is never what was predicted.
Travel data is a series of near misses: booking lifecycles, channel conflict, and a demand forecast that is wrong in the direction of the trip not taken.
Read more → - Transportation
Fleet, freight, and the network that moves — the empty truck is the enemy.
Transportation is the data of movement, where the truck itself is the smallest unused slot, and the network is ahead of the metric that spoke.
Read more →
Partnerships
Platforms we build on
Five partnerships, each earning its place. What matters is what we do with them.
Models
Anthropic
Claude is the default model in our assistant and document work. Partner status gives us early access to capability and safety guidance we pass through to clients.
Lakehouse
Databricks
Our primary platform for lakehouse architecture, pipeline orchestration, and putting governed data next to model workloads.
Warehouse
Snowflake
Where we build warehouse and semantic layers for organisations already standardised on it.
Infrastructure
AWS
Most systems we build run in the client's own AWS account, under their controls and their billing.
Cloud & workspace
Google
Cloud infrastructure and workspace integration for teams whose day runs inside Google.
Insights
Notes from the work
Jul 22, 2026 · 2 min
Most failed AI pilots are data failures with better branding
Six pilots that stalled, and what was actually wrong in each. In five of them the model was fine.
Jun 18, 2026 · 2 min
Write the evaluation set before you write the prompt
Forty graded cases, drawn from real inputs, cost about a day. Skipping them costs about six weeks.
Tell us what you're trying to do.
A paragraph is plenty. We will tell you whether it is a job for AI, a job for a pipeline, or not a job at all.