Engineering for systems you cannot switch off
A vendor system with no API. A production line that never stops. A residency clause that rules out every hosted product. Those constraints are the job, not an obstacle to it.
- 100+
- Websites built
- 200+
- Campaigns managed
- 7
- Years running
Three industries, and the thing each one keeps hitting
The six services do not change between them. What changes is which part of the job comes off your desk first, and that is worth reading in your own words.
Forex and trading
Cost per lead is known and cost per funded account is not
Read the page
E-commerce
Mobile is most of the traffic and the worst of the conversion
Read the page
News and media
More than a quarter of engineering time is unplanned
Read the pageNumbers we published, and where each one was read
Client names are withheld under NDA. The numbers are not: measured over a stated window, on the client's own tooling.
Logistics, UAE. Datadog, 90 days
Professional services, US. CRM log
Education, UAE. 12 months vs prior 12
Twenty eight full engagements across them, in seven client countries. Aggregate only: the disclosure level of every case is published on the case itself.
Six, not sixty. Most engagements start in one and pull in a second when the constraint turns out to live there.
AI product engineering for companies that already have systems
Most of our work starts where a greenfield build ends. The interesting decisions are not about what to build, they are about what the existing system will and will not allow.
We are a small senior team. Every engagement has a named baseline before it has a plan, and a named internal owner before it ends.
Operating principles
- 01
Measure before you build. No baseline, no project.
- 02
Write down the option you rejected and why.
- 03
The smallest change that moves the number wins.
- 04
If your team cannot run it without us, we failed.
Six services, one engineering standard
Most engagements start in one and pull in a second when the constraint turns out to live there.
Client names withheld, numbers are not
Every engagement below is real and anonymized to the level the contract allows. Figures carry the measurement window that produced them.
A tracking portal that stopped queueing behind itself
Peak hour timeouts on a shipment tracker, caused by a vendor system nobody was allowed to touch.
- p95 response
- -89%
- Throughput
- +290%
- Tickets
- -54%
CS-2025-01
The chatbot we talked them out of building
Losing deals to response time, with every proposal assembled by hand from prior engagements.
- Turnaround
- -96%
- Hours returned
- +58
- Win rate
- +11%
CS-2025-07
Deleting nine apps beat rebuilding the store
A Shopify storefront carrying fourteen apps, losing mobile buyers before the page finished painting.
- Mobile LCP
- -69%
- CLS
- -89%
- Conversion
- +28%
CS-2025-03
Where the twenty eight engagements sit
Counts include work we cannot publish. Each row is a sector we have shipped in, never one we would merely consider.
| Sector | Engagements | Pillars used |
|---|---|---|
| Logistics | 4 | AI, Web |
| Fintech | 5 | AI, Growth |
| E-commerce | 4 | Web, SEO |
| Healthcare | 3 | AI, Growth |
| Education | 3 | SEO, Paid |
| Manufacturing | 2 | AI |
| Professional services | 3 | AI, Growth |
| Real estate | 1 | Web, Paid |
| Media | 1 | Web, SEO |
| Retail | 2 | Growth, Paid |
Four phases, and you can stop after any of them
Assess
Measure the baseline before proposing anything. If there is no baseline we build one first.
Decide
Target design with the rejected alternatives written down, so the choice is auditable later.
Build
Reversible increments behind flags. Every deploy measured against the baseline.
Hand over
Runbook, named internal owner, and the metrics your team watches after we leave.
We publish our own numbers because we sell them
An agency that sells performance and ships a slow site has answered the question already. Every figure is a Lighthouse run against this build, including the mobile one that is not flattering.
Lighthouse desktop, every route tested
Lighthouse mobile, throttled 4G
WCAG 2.2 AA, zero failing audits
CLS on every route tested
Slowest route measured
Throttled 4G. The React runtime is the floor
Measured 21 August 2026 against the production build. Three runs per profile, median reported.
Who this works for, and who it does not
Saying no to the wrong engagement is the cheapest thing either of us can do. Here is where we are honest about it.
A good fit
- You already have systems, customers and revenue to protect
- Someone internally can own the change after we leave
- There is a number that should move, and you know which one
- Procurement, security and legal are part of the plan, not a surprise
Not a fit
- You want a demo for a board meeting rather than a system in production
- The deadline is shorter than the shortest honest path to the outcome
- Success is defined as shipping rather than as a measured change
- The work has no owner with authority to approve a decision
Start with a measurement, not a meeting
Buy the two week assessment and bring a real report to the first call, or book 45 minutes and bring the constraint. Either way you arrive with something concrete.