Build
AI agents that do real work
LLM applications, retrieval systems and agents that operate inside your business rather than beside it — with evaluations, so you can tell whether they still work next quarter.
Sound familiar?
You have a prototype that demos well and has never been trusted with a real transaction.
Nobody can tell you whether last week's prompt change made the system better or worse.
The interesting work is locked behind your operational systems, and the AI has no safe way to reach it.
What we do
Capabilities
- Agents inside your ERP
- Invoice triage, purchase-order matching, close assistance and procurement workflows that read and write against your real system of record.
- Retrieval over your own data
- RAG that works on your documents and records, with the chunking, ranking and permissioning that decide whether it is useful or merely plausible.
- Evaluations
- Test sets, regression suites and scoring, so a model or prompt change is a measurable decision rather than a vibe.
- Human-in-the-loop design
- Deciding what the agent may do alone, what needs approval, and how a human takes over. Mostly a process question, not a model question.
- Guardrails and observability
- Input validation, output constraints, cost controls, tracing and audit logs — the parts that separate a production system from a demo.
- Model selection and cost
- Picking models against your accuracy and latency budget, and reducing cost without silently degrading quality.
Outcomes
What changes
Before
A pilot demos well and never reaches anybody's daily work.
After
The agent runs inside the process, against the real system of record, with a person approving what matters.
Before
Nobody can say whether the model is getting better or worse.
After
It is scored against an evaluation set on every change, so regressions are caught before release.
Before
The agent's mistakes are found by the person they were made for.
After
Low-confidence cases are routed for review before they reach anyone.
How it works
The engagement
- 01
Use-case selection
1 week
Choosing a workflow where the volume is high, the rules are knowable and a mistake is recoverable. Most failed AI projects pick wrong here.
- 02
Evaluation harness first
1–2 weeks
We build the test set before the feature. Without it you cannot tell improvement from regression, and every later decision is guesswork.
- 03
Build
4–8 weeks
The agent, its integrations and its guardrails, scored against the harness on every change.
- 04
Supervised rollout
4 weeks
Live but human-approved, until the numbers justify loosening the leash. Autonomy is earned, not configured.
Go deeper
The playbooks set out how these processes run, step by step — what is standard, what gets configured, and what changes.
Stack
- Claude
- OpenAI
- LangGraph
- pgvector
- Python
- TypeScript
- Temporal
Questions
The things people actually ask
- Where should we start?
- A high-volume, rule-shaped workflow where a mistake is caught cheaply — invoice coding, PO matching, document triage. Not customer-facing chat, which is the most visible and least forgiving place to learn.
- How do you know it works?
- We build the evaluation set before the feature and score every change against it. If someone cannot show you a number, they are telling you it feels better.
- Is our data safe?
- Enterprise API tiers from the major providers do not train on your data by default. The larger risk is usually your own retrieval layer leaking records across permission boundaries, which is a design problem we handle explicitly.
- Should we fine-tune a model?
- Almost never first. Better retrieval, better prompts and better evaluation beat fine-tuning for most business workflows, at a fraction of the cost and without pinning you to a model version.
- What does this cost to run?
- Inference cost scales with volume and is usually smaller than teams expect — often less than the labour of the workflow it replaces. We model it during discovery so there are no surprises at scale.
- What if the model changes underneath us?
- That is exactly what the evaluation harness is for. When a provider ships a new version you re-run the suite and find out in an hour instead of discovering it through a customer complaint.
Tell us what’s breaking.
Thirty minutes, no deck. We’ll tell you whether this is the right place to start.

