Approach
Approach
Four phases, each with a decision point. The point of the structure is simple: the real build only starts once cheap, fast work has proven the idea is worth it and the data is there.
-
Discovery
1–2 weeks
We find the assumption most likely to sink the project and test it first — usually data access, a third-party dependency, or the hardest edge case. Everything else waits.
Outputs
- Written viability verdict: go, no-go, or go-with-conditions
- Audit of your data and systems
- Scoped plan and budget for the prototype phase
-
Prototype
1–3 weeks
A thin, end-to-end version running on your real data and real edge cases — not a curated demo. Its job is to produce evidence, including uncomfortable evidence, before you commit to the full build.
Outputs
- Working prototype in your environment
- Honest results: where it works, where it fails, and why
- Running-cost projections at your real volume
- Go/no-go recommendation for the full build
-
Build
2–10 weeks
The real thing, built with AI-assisted engineering so it lands in weeks rather than quarters. Security, monitoring, tests, and recovery paths are part of the definition of done from the first commit.
Outputs
- Production system in your own accounts and repositories
- Automated tests and evaluation gates
- Threat model and applied hardening
- Documentation and operational runbook
-
Handover
1–2 weeks
You run it and we watch, then we step back. Handover has its own deliverables and acceptance criteria because a system only we can operate is not finished.
Outputs
- Walk-through sessions with whoever will own it
- Plain-language playbook for day-to-day operation and what to do when something breaks
- Agreed checkpoints for follow-up review
Principles
How we make decisions when nobody is watching
Say no early
The cheapest failure is the one found in discovery. We would rather return a no-go verdict in week two than a polished excuse in month six.
Real data or nothing
Prototypes run on your actual data, volumes, and edge cases. Curated demo sets prove nothing except that demos are easy.
Cheaper, never careless
AI-assisted engineering cuts the cost and time of building. It never cuts the security review, the tests, or the documentation.
Server-side decisions
Anything that matters — permissions, limits, spend controls, validation — lives where users cannot bypass it. The browser is never trusted.
You own everything
Code, prompts, automations, infrastructure, and documentation land in your accounts. No proprietary layer, no lock-in, no invoice for leaving.
Write it down
Every non-obvious decision gets written down. Whoever inherits this in two years should never have to guess why.
Ready for a discovery phase?
It is short, fixed-scope, and ends in a written verdict either way. Tell us what you are weighing up.