AI implementation audit / $7,500 / 2 weeks
Companies fall behind on AI for ordinary reasons.
I am a working software engineer. I build retrieval systems over private corpora and agent workflows that run in production. Lapointe Labs is one fixed-scope engagement, for companies and for agencies: a two-week audit that finds those ordinary reasons and puts numbers on them.
- Fee$7,500 fixed
- Duration2 weeks
- Report15–25 pages
- Evidencemeasured before & after
A model demo takes an afternoon. What takes months is everything after it: the project that stalled, the team whose workflow never caught up, the business that has not started at all. The reasons are ordinary — nobody owns it, nobody measures it, nobody wrote down what done looks like. Those problems are unglamorous, they cross teams, and they are usually nobody’s job. Making it somebody’s job is the work.
01 / The offer
AI Implementation Audit
Fixed scope, fixed price, quoted before the work starts.
What I look at
Retrieval path. How documents are ingested, chunked, embedded and refreshed, and what the index actually returns for twenty real questions taken from your users.
Ranking. How often the correct passage is retrieved but ranked too low to be used. That is a different repair from a document the index never had.
Evaluation. Whether anyone could detect a regression today. I build a small labelled eval set from your own queries and score the current system against it.
Agent and tool boundaries. What the model is permitted to call, what runs without a human in the path, and what happens on a timeout, a partial failure, or a third retry.
Context assembly. What actually lands in the context window on a real query, how the token budget is spent, and what is being truncated before the model sees it.
Ownership. Who can change a prompt, who can ship an index rebuild, how long that takes today, and what each of those people is currently waiting on.
Deliverable
A written report of 15 to 25 pages, plus the work behind it.
The report ranks what is blocking the project and names specific defects with file and line references wherever I have repository access. Each item carries a proposed fix, a rough estimate, and a priority, and the sequenced plan is marked to show which parts your team can do and which parts need outside hands.
Shipped with it: the labelled eval set and the scoring scripts, committed to your repo and runnable in your CI if you want them there, plus a reproduction for each failure I found. Then two hours with your engineers to go through all of it, recorded so the people who miss the call can watch it.
The report and the eval set are yours.
Timeline
Two weeks, from access granted to walkthrough.
What this is not
Acting on any of it does not require hiring me again. If an audit is the wrong thing to buy, I say so on the call and tell you what I would do instead.
How it runs
Call, 30 minutes
You describe what you built and where it stopped. I ask about the corpus, the query volume, and what a wrong answer costs you.
Access, week one
Read access to the repo, the index, and a sample of real queries and real documents. Forty-five minutes each with the two or three people closest to the system. A copy of the questions real users are asking, however messy the list is. Your standard NDA is fine.
Measure, weeks one and two
I build the eval set, score the current system, run it against those real questions one at a time, and break it on purpose. Where I cannot see what happened, I add the logging and tracing that is missing. Findings land in a shared doc as I get them, with a short written checkpoint at the end of week one, so nothing in the report arrives as a surprise.
Report and walkthrough, end of week two
You get the report, the eval set and scoring scripts, and the sequenced plan, plus two weeks of email while your team starts working from it. Then we decide whether you implement it, I implement it, or nobody does.
02 / Price context
The number you are actually comparing it to
$250,000yr 1, approx.
$7,500fixed
approx. 3%of that first-year figure
These are approximate market figures for orientation, not a study. Salary plus payroll tax, benefits and equity for a senior engineer in a US metro lands somewhere near a quarter of a million dollars in the first year, before any recruiting fee, and searches at that level are commonly quoted at three to six months before anyone starts. Bars are drawn to scale from the two figures above. The audit does not replace the hire. It tells you what to hire for.
03 / Who does the work
One engineer, named, on the call
I am Marc Lapointe. I am a working software engineer: retrieval systems over private corpora and agent workflows that serve real users, in production, right now. This is what I do every day, and have for over seven years. Every opinion on this page comes from that work — not from a conference circuit, and not from a playbook written when the tooling was different.
I take a small number of engagements at a time, on purpose. The person on the call is the person in the repo: no handoff, no account layer, no junior associate learning on your budget. When I commit to your two weeks, they are yours.
I have no case studies, and I am not going to borrow anyone else’s. The proof is the work I do now: production retrieval and agent systems, serving real users. On the call, ask me anything — chunking strategy, eval design, retry semantics, reranking, cost per request. Judge the answers.
04 / For agencies
Delivery capacity for work you have already sold
How it works
I work as white-label AI delivery capacity under your name, in your repo, on your standup. Retrieval, agents, evaluation infrastructure, built to your standards and documented so your team owns it after handoff.
Your client stays yours
I do not contact your client, their name does not appear on my site or anywhere else, your client stays your client as a term in the contract, and I will sign the non-solicit you already use before I see a line of code.
Capacity
I take a defined number of delivery days each month, and I will tell you exactly what that number is before you commit to a date. Booked days are yours: no bench, no sharing, no quiet reassignments.
Rates
Rates depend on what you have sold and how fast you need it, so we settle those on a call. Bring me in before you scope it if you can.
05 / Next step
Tell me where it stopped.
If your AI work has stopped moving, book thirty minutes. Tell me what you were trying to ship and where it stopped, and I will tell you on that call whether the audit is worth $7,500 to you.
Book a 30-minute call