Theme
KeelwrightSystems
● Live
Clip
KeynoteThe cost layer of AIMain stage
Cost per resolved ticket$0.030
Inference$0.019
Retrieval$0.008
Retries$0.002
Logging$0.002

30 min · Free · Video call

Step 1 of 4

What should we look at?

Pick the closest fit. We'll refine it on the call.

Step 2 of 4

Pick a time

Times in Gaborone (UTC+2), with your local time underneath.

Step 3 of 4

Your details

Step 4 of 4

Check your request

Your request is ready

Last step: send it to Quinton on LinkedIn. He confirms the slot and sends a video link. Nothing has been sent yet.

Open LinkedIn ↗

( Measure first )
Since2026

Keelwright Systems. DevOps, platform and AI cost consultancy, Gaborone, Botswana. Working UK and EU hours.

K( Measure first )

( A practice built on one precise point )

Every strong system depends on a part nobody sees: the keel. Keelwright puts a number on AI features, platform debt and prompt logs, then builds the fix inside your stack, so finance, legal and engineering read the same page.

Founded by Quinton JosephsBW / UTC+2

Cost, risk and delivery, measured where they happen: in your repos, your billing and your traces.

SCROLL
( Practice )
AI cost / Platform TCO
Data protection / Builds
01

We price a resolved ticket end to end, with retrieval, retries and logging included, then cut it where it is fat.

( Cross-discipline infrastructure )

Four lines of work, each with a defined end and a deliverable you can hand to finance, legal or the next engineer. Fixed fee, quoted after a 30-minute scoping call.

( AI unit economics )

2W
AI cost
line review

We price a resolved ticket, a generated document or a classified record end to end, then cut it where it is fat.

  • Cost per outcome, up to 3 features
  • Model routing and retrieval depth
  • Per-tenant budget caps
  • One-page summary for finance

( Data protection )

1W
Prompt
log audit

The prompt is the new database row. We inventory every sink a prompt touches and set one rule per sink.

  • Inventory of every sink
  • Allow, block or review rules
  • Redaction and retention PRs
  • Evidence pack for your DPO

( Platform TCO )

3W
Tech-debt
register

Tech debt behaves like a loan. We price the monthly interest and pay off the most expensive piece first.

  • Owners and interest estimates
  • Platform cost model
  • Quarterly refinance plan
  • CFO-ready summary

( Hands-on builds )

4–6W
Platform
build sprint

Multi-region Terraform, Kubernetes ingress, CI/CD for LLM containers, and brittle bash rewritten in Python or Go.

  • Merged into your repos
  • Runbooks and tests
  • Handover sessions
  • Two weeks of follow-up
( Who it is for )( UK / EU / Southern Africa )

The Sectors

Scale-ups and mid-size teams in the UK, EU and Southern Africa, plus the agencies and cloud partners who serve them.

( Selected posts )Follow on LinkedIn →
W03

Pricing tech debt like a loan: principal, interest, default

Platform TCO+

This is one of the more useful ways to price tech debt I've seen
(borrowed from how the best CFOs I've worked with read engineering budgets)

A few notes on how it works:

Don't ask "how much debt do we have?"
Ask what it costs you every month to keep it.

Debt behaves like a loan:
Principal → Interest → Default
The principal is the rewrite. The interest is the toil, incidents and slow deploys you pay while you wait.

Interest is the number finance understands
If a 40-person team spends a fifth of its hours on debt service, that's about eight engineers' salaries every year. Nobody approves that line on purpose.

Default risk matters more than size
A small module that blocks every release is worse than a large one nobody touches.

Refinance in slices
Pay down the highest-interest piece each quarter and measure the drop in interest, the way you would with any loan.

One practice worth copying:
every item on the debt register gets a named owner and a monthly interest estimate, reviewed alongside cloud spend.

As an engineer, I want to rewrite the whole thing.
As a founder, I know the math says pay off the most expensive interest first.

How do you put tech debt in front of your CFO today: as a backlog, or as a monthly cost?

W02

PII controls must cover every place a prompt is logged

Data protection+

PII controls can't stop at the database.
For LLM pipelines, they have to happen everywhere a prompt gets written down.

Take something simple like logging a request.
Your service already has permission to do it, and your observability stack will keep it for weeks by default.

Logging token counts and latency to Datadog? Probably fine.

Logging the full prompt, including a passport number a customer pasted into support chat? Definitely not.

Logging the retrieved chunks so you can debug bad answers? Maybe, with redaction and a shorter retention window.

This gets harder as tracing gets easier.
One environment variable in a LangChain or OpenTelemetry setup can start shipping full prompts to a third-party backend, and nobody opens a PR for it.

When I set this up for clients, I use one rule per log sink:

  • Allow: metadata only (model, tokens, latency, trace ID)
  • Block: raw prompts and completions in any shared sink
  • Review: retrieved context, redacted with Presidio and kept for a short, fixed window in a separate bucket

Log volume drops sharply.
The next DPIA gets much shorter too, because you can point to exactly where every prompt lives.

The prompt is the new database row.
Treat every sink that touches it like one.

Where do your full prompts end up today, and who outside the team can read them?

W01

The CFO's unit-economics review is AI's toughest proving ground

AI cost+

The CFO's spreadsheet may be enterprise AI's toughest proving ground.

An AI feature can pass every eval and still get shelved if nobody can say what one request costs.

When consulting for clients, I often see this play out in the same meeting.
The demo lands, finance asks for the cost per resolved ticket, and the room goes quiet.

I've watched support assistants with 90%+ accuracy sit paused for weeks.
Nobody could forecast the bill past pilot volume.

What gets them back into production is boring work on four lines of the stack:

  • Cost per outcome. Price a resolved ticket end to end, with retrieval, retries and logging included, and it often lands in cents rather than the dollars people fear.
  • Model routing by task. Most routine traffic can move to a smaller model, with the large one kept for escalations.
  • Hard budget caps per tenant. A runaway prompt loop hits a ceiling before it hits the invoice.
  • Retrieval depth tuned against quality. Many teams retrieve far more chunks than their answers need, and cutting top-k is the cheapest test you can run.

The same test applies well beyond support bots.

If you can't put an AI feature on a unit-economics line, you don't have a product yet.
You have an experiment with a company card attached.

What unit do you use to price your AI features today: tokens, requests, or business outcomes?