Technical assessment
Your stack, your team, the plan as it stands.
- Read the code and the design so far
- Map the systems the build must touch
- Find the parts with the most risk
- Agree who does what on both sides
- Set the tests the build must pass
01 / 04
Your team can build it, they have never built this
Retrieval that returns the wrong passage, an agent that loops, a connector that breaks the ERP. Each costs a week if nobody in the room has seen it before, and the engineers who could learn it are busy.
CH.01 · The problem
Your team can build it. They have never built this.
The first one is the hard one
Retrieval that returns the wrong passage, an agent that loops, a model that behaves differently at load. Each is a week lost if nobody in the room has seen it before.
Your systems are the hard part
The model is the easy piece. Getting it to read the ERP, respect the permissions and write back to the CRM without breaking anything is where projects stall.
Nobody can be spared
The engineers who could learn this are the ones keeping everything else running. A project that needs them full-time for six months does not start.
Our answer
Pair our engineers with yours, on your code
We join the build: design reviews, pairing, the hard integrations, the tests that catch what demos hide. Your team writes most of the code and keeps all of it. We leave when it runs without us.
CH.02 · What we build
How we work in your build
Three parts: know the ground, build alongside, prove it works.
Engineers who have shipped thisInside the buildPatterns that hold upTested before anyone trusts it
Building together
Your engineers with ours, on the same code.
- Pair on the parts your team has not built before
- Take the hardest integrations ourselves
- Review every pull request against the design
- Unblock daily; nobody waits a week for an answer
- Write the runbook as the system takes shape
Quality assurance
Numbers before release.
- Build the evaluation set with the domain owner
- Test each connector against the live system in read-only mode
- Load-test at the expected peak
- Run a week in shadow mode beside the current process
- Report results against the agreed targets
CH.03 · How it runs
How it runs
Four phases, in your repository.
- 01Phase 1
Assessment
Code, systems, plan, risks.
- Review the design and the code so far
- Map the systems and the access needed
- List the risks and how each will be retired
- 02Phase 2
Planning
Phases, owners, tests.
- Split the build into phases with deliverables
- Assign each part to your team or ours
- Schedule the pairing sessions
- 03Phase 3
Build
Pairing, reviews, integrations.
- Pair on the new patterns
- Build the hardest integrations
- Review pull requests daily
- 04Phase 4
Validation and handover
Prove it, then leave.
- Run the evaluation set and report
- Load-test and fix what it finds
- Shadow mode beside the current process
CH.04 · What changes
What changes
Your team ships it, learns it, and owns it.
Weeks instead of a lost quarter
The problems that stall a first AI build have been seen before. Your team hits them with someone in the room who knows the way out.
The code stays yours
Your repository, your standards, your engineers on every pull request. Nothing is handed over because nothing left.
Your team can build the next one
Pairing is the training. After one build alongside us, your engineers know the patterns, the tests and the traps.
| Side by side | On your own | With us in the build |
|---|---|---|
| Speed | Every problem is new | Most problems have been solved before |
| Quality | Learned from incidents | Evaluation, integration and load tests before release |
| Risk | Found in production | Listed in assessment, retired in the build |
| People needed | Your best engineers, full-time | Your engineers part-time, ours for the hard parts |
| Afterwards | One system, hard-won | One system and a team that can build the next |
CH.05 · Questions
Questions
The ones we build ourselves: retrieval over company documents, agentic workflows, fine-tuned and local models, and the integrations with ERP, CRM, mail and ticketing around them. Machine learning projects such as forecasting and churn as well. If the project is outside that, we say so in the first call.
In the build, in your repository. Pairing sessions on the new parts, pull request reviews every day, and the hardest integrations written by us. The share of code your team writes grows through the project, by design, so that at handover they own all of it.
An evaluation set built with the domain owner for the model's answers. Integration tests for every connector, run against the live system in read-only mode first. A load test at the expected peak. A week in shadow mode beside the current process. Each produces a number reported against a target agreed in assessment.
Through their APIs, under the permissions they already enforce, with the least access the task needs. Writes are wrapped in checks the system cannot skip and tested in read-only mode before they are enabled. Your platform team reviews every connection.
More in AI Consulting
Back to AI ConsultingThe map of the practice
Start
Bring us the problem nobody has cracked yet.
We are a small team of senior specialists. We pick the right model and the right layer, and we build the least machinery that does the job. You get a call with an engineer, not a sales deck.