Industries
Six sectors, six different reasons the obvious build fails
Financial services, healthcare, retail, manufacturing, education and logistics. The pipelines look alike in all six — what differs is the one thing that is not allowed to happen, and that is what decides the architecture.
Coverage
- Sectors
- 6
- Applications mapped
- 42
- Constraints documented
- 30
A capability map, not a client list. We name no client, by policy.
Why sectors
We do not run six stacks. We design around six failure surfaces.
A retrieval pipeline in a hospital and a retrieval pipeline in a warehouse are the same seven components. The difference is what happens when one of them is wrong.
Most of the engineering transfers. Vector search, an evaluation harness, a feature store, a deploy pipeline, an on-call rotation — none of that is sector knowledge. A team that has built one of them can build the next one, and any firm claiming otherwise is selling a vertical badge.
The transferable part is not where programmes fail. They fail on the constraint discovered in month three — the model endpoint that turned out to be a cross-border data transfer, the plant network that will never accept an inbound session, the search latency budget already spent before anyone proposed a reranker. By then the architecture has been chosen.
So we organise by constraint rather than by vertical marketing. Each sector page names the forces reshaping it, the applications worth building now, and — at length — the constraints that decide the design. The constraints section is the one worth your time.
If a sector page reads like it could be any sector, we wrote it wrong.
Where we work
Six sectors we know the failure modes in.
Each one lists the applications we would build today, the constraints we design around, and the blueprints behind them.
- 01
Financial Services
Credit, fraud and compliance systems that can explain themselves to a regulator years later
What we build here
- Onboarding document intelligence
- AML alert triage and narrative drafting
- Credit decisioning with real reason codes
- 02
Healthcare & Life Sciences
Clinical and operational systems where being wrong is expensive — built with the citation attached
What we build here
- Ambient documentation that produces structured facts, not just prose
- Prior authorisation evidence assembly
- 03
Retail & Commerce
Catalogue, search and demand systems that hold their budgets on the biggest sale day of the year
What we build here
- Catalogue enrichment and attribute extraction
- Hybrid retrieval, query understanding and ranking
- 04
Manufacturing & Supply Chain
Plant-floor signal turned into planning decisions — without ever touching the safety layer
What we build here
- OEE and downtime attribution that survives an argument
- Visual inspection running at line rate
- A shop-floor copilot over SOPs and maintenance history
- 05
Education & EdTech
Scoring, tutoring and credentialing that a candidate can dispute — and you can defend
What we build here
- Rubric-aligned scoring for written responses
- Speaking and pronunciation scoring
- 06
Logistics & Mobility
Routing, ETA and exception systems that hold up when the network is noisy and the margin is thin
What we build here
- Address normalisation and rooftop geocoding
- ETA that ships a distribution, not a point
The deciding factor
One constraint dominates each sector.
Not the only one, but the one that shapes everything downstream of it. Read the middle column as the reason the right-hand column is not a preference.
| Sector | The binding constraint | What it forces in the architecture |
|---|---|---|
| Financial ServicesExplainability | Explainability is not a feature you add later | An interpretable model of record, reason codes produced by the same computation as the score, and a challenger running in shadow from day one. |
| Healthcare & Life SciencesMinimisation | PHI minimisation has to survive free text | Send the span a task needs rather than the whole record, pin inference to zero-retention in-region endpoints, and document the de-identification decision. |
| Retail & CommercePeak load | Peak is not a bigger Tuesday | A system designed to shed — precomputed ranking for the head of the catalogue, a BM25-only degrade path, and a shedding order agreed before sale night. |
| Manufacturing & Supply ChainSegmentation | The OT/IT boundary is not negotiable, and should not be | A read path that terminates in the plant DMZ and publishes outward from a broker the plant owns. Nothing we build writes to a controller. |
| Education & EdTechDisputed scores | A disputed score is the real specification | Model version, prompt, rubric revision and exemplars pinned into an immutable scoring artefact, with a documented route to a human re-mark. |
| Logistics & MobilityLocation error | The location is wrong twice over | Map-matching over the road graph with explicit gap handling, and no component anywhere that treats a raw GPS ping as truth. |
Each sector page states four more constraints beyond the one listed here.
Next step
Tell us the constraint, not the use case.
A 45-minute working session. We'll tell you what we'd build, what we'd not build, and roughly what it costs. No deck.