I make RAG and agents work on the systems your client already has.
The model is maybe 20% of the cost. The integration is the other 80% — and that is the part that goes wrong. I build retrieval and agents inside your client's existing Postgres estate and monolith, with a measured recall floor and the failures published. White-label, fixed price, source handed over. I never appear in your proposal.
UTC+1 — overlaps EU business hours; 13:00–17:00 UTC US-East overlap
Why brownfield, and not greenfield
53%
of organisations cite legacy integration complexity as their top barrier to scaling AI.
9%
of leaders say they fully understand their own applications' dependencies.
81%
say technical debt is already constraining their AI success.
HFS/Genpact Data Modernization and AI 2026 (n=48) · Kyndryl Enterprise IT Modernization Trends, October 2026 (n=2,000) · CAST via IBM brownfield modernization study, February 2026.
Almost nobody can hire for this. The person who understands both the ML side and the legacy estate well enough to design the integration barely exists as a defined job — which is exactly why an agency subcontracts it.
What I actually do
Retrieval against your client's own corpus
Extraction, sentence-aware chunking, embeddings, vector search with an HNSW index, answers carrying their source. Built against their schema and their permission model, not a clean-room demo database.
Tool-calling agents that terminate
LangGraph state machines: conditional routing, bounded loops, tool arguments validated against a schema before execution, and human handoff that carries the whole conversation rather than a summary of it.
Writes behind approval gates, with an audit trail
An agent that creates a record is a different risk class from one that answers a question. Every write path goes behind a human approval queue and lands in an audit log. This is the part most demos skip and most production incidents are about.
Cost and latency you can defend
Token usage and p50/p95 latency recorded from the first milestone, not retrofitted. Where the engagement is cost-shaped, the fee can be indexed to a verified reduction reconciled against the provider invoice — with a floor, so the downside is bounded.
How an engagement runs
01
NDA before anything moves
Mutual, signed before the first substantive call. If your client requires one, I use yours.
02
I look at the estate before I quote
The schema, the data hygiene, the permission model, what already exists. A price quoted without looking at the system is a guess, and you'd be paying for the guess.
03
Fixed price, and the number comes first
One figure, tied to a measurable outcome — a retrieval recall floor, or a token-cost reduction reconciled to the provider invoice. Not an hourly meter.
04
We build into what exists
Your client's Postgres, their ORM, their auth. No rebuild of what works, no new platform subscription, no parallel system nobody adopts.
05
Measured before handover
Retrieval scored against a goldset built from their own corpus. Refusal rate on out-of-scope questions. p50/p95 latency and cost per conversation.
06
Source and infrastructure transferred to your accounts
Your repo or mine, your cloud accounts, your keys. Not held in mine and rented back.
What you get, and what I don't do
What you get
- Production code with documentation — not a zip file and a wish
- A committed goldset built from your client's documents, in your repository, so you can rerun it
- A measured recall floor and a published refusal rate, which is what your client's risk review will ask for
- Severity-tagged eval data that gates deploys, so the AI cannot silently regress in production
- A handover call with your team, recorded
- A 30-day defect window
What I don't do
- I don't appear in your proposal. Your name is on it.
- I don't talk to your client unless you ask me to be in the room.
- I don't hold your source, your infrastructure accounts or your keys after handover.
- I don't start engagements without written scope and acceptance criteria.
- I don't quote an hourly meter when the work has a measurable outcome.
Proof you can check before you reply
You are going to verify this, because you are the one who gets blamed if it is wrong. So here is the part most suppliers cannot show you.
The benchmark
232 labelled queries against a 45-chunk knowledge base. Full threshold sweep, committed goldset, one command re-runs it.
The unflattering number
At the F1-optimal threshold the vector path scores 93.4% F1 and a 76.9% false-positive rate. It answers 20 of 26 out-of-scope queries. That number is on the page on purpose.
A side-effecting agent
Reads WhatsApp, checks real stock, creates orders — behind human approval gates with an audit trail. Live, no signup.
The systems themselves
Five, each labelled for what it actually is, with limitations stated rather than buried.
Two named client references exist. Both were website builds, not AI systems, and the homepage says so rather than dressing them up. What they establish is remote delivery — on schedule, twice, for real businesses in the UK and Cyprus.
Rate
From $8,000 per project
Or $12,000 / month. Minimum three weeks. Deposit under 50%, milestone payments, source and infrastructure transferred to your accounts at handover.
On a cost-shaped engagement I will index the fee to a verified reduction and reconcile it against your provider invoice, with a floor so the downside is bounded. On a quality-shaped engagement the price is fixed against a recall floor we agree up front. In both cases the number is the deliverable, not the hours.
This is not for you if…
- You want a greenfield demo on a clean schema. That's a smaller job than I'm priced for, and there are cheaper people who do it well.
- You want me to speak to your client directly. This is a delivery relationship, not an agency-of-record one.
- You need it live in under two weeks. That's a capacity problem, not a scoping problem, and I'd rather tell you now.
- The budget is under $8,000. At that level you want a scoped build on its own, not a subcontracting relationship.
Published rather than discovered. A fixed price only means anything if the edge of it is written down.
Send me a client with a legacy estate and an AI feature that isn't behaving.
I'll tell you in writing whether it's fixable, and what the number is. If the answer is that you don't need a build, you'll get that instead — and you keep the report either way.