← fractionalmin.com
Impact · Selected engagement

One week to make the team twice as fast. Then nine times the output, without me.

JENGAI had real customers and a round closing. In a single week alongside the team, I doubled how fast each engineer shipped and left them an AI delivery system they own. It kept compounding after that. By July, eight engineers were shipping 537 deliverables a month against 60 in May, and my own commit count had fallen to roughly zero.

What began as Weeklong Agentic Workflows Training became an ongoing engagement. The week proved it, then Shayne kept me on to clear the performance ceiling and build real AI into the product itself.

60 → 537
Deliverables shipped per month
May to July. Features, user stories and fixes, not commits.
100%
Of features carry a PRD, a TDD and a checklist
Not a sample. All of them. It is where their estimates now come from.
13m → 15s
A core read, fixed
One of several order of magnitude wins the team could not crack alone.
~0
Commits from me now
Eight engineers run the whole system without me.
The velocity, measured honestly
Commits per engineer, per month
The team, not counting my own output, normalized for headcount, so this is each person shipping more rather than more people shipping. July is the mean; the median engineer shipped 299. Commit count on its own is a weak measure, and this method deliberately produces smaller commits, so read it alongside the deliverables above.
I joined mid May
53
70
29
57
42
93
241
Jan
Feb
Mar
Apr
May
Jun
Jul
Dashed line: the average before the install, about 52 commits per engineer per month.
What the engagement actually was
the weeklong training

Made the team twice as fast, and left them owning it

One week working alongside them. Each engineer came out shipping about twice as fast, and it held, because I installed a delivery system they run themselves. Idea to spec to review to ship, with automated senior review built in before a human ever looks.

ongoing · remediation

Cleared the performance ceiling

Instrument first, build a safety net, then fix. A core read went from 13 minutes to about 15 seconds. A common delete went from hundreds of thousands of database reads to a handful. The class of problem raw AI gets confidently wrong.

ongoing · the system

Built a library of reusable skills and agents

Continuous integration and deployment, async infrastructure, a test and safety harness, observability, and a set of skills and agents the team keeps using. The foundation every feature now ships on.

Where the growth actually went

Output grew about five times faster than defects did.

This is the number that matters more than the velocity. If a team ships five times more and breaks five times more, nothing was gained. From May to July, feature work grew 20x while defect work grew 3.9x. More capacity went into new product, not into repair.

Feature work
20x
Defect work
3.9x

July's 537 deliverables were 419 feature tasks, 60 user stories and 58 fixes. In May those were 21, 24 and 15.

ongoing · the bigger ask

Then the part most teams do not know how to attempt: I built real AI into their product.

Not developers using AI to write code faster. AI built into the product itself. An assistant that can safely propose, preview, and undo real changes to live customer data, backed by full change history. Most companies this size cannot attempt it, and most do not yet know where to start. This is the work I am finishing for JENGAI now, and it is the capability I most want to build for the next team.

The math the board actually cared about: value in, value out.

The lift was not a one week bump. It compounded for three months after the install week, and it kept compounding once I stopped writing code, because the team owns the system rather than renting it from me.

Velocity, first month~2x
Commits per engineer by July~4.6x
Median engineer, July299
Deliverables, May to July60 → 537
Feature work vs defects20x vs 3.9x
Commits from me now~0
Who owns it nowthe team
Risk closed along the way
security

Closed a live security exposure

Found it, closed it, and moved every secret into a managed vault.

reliability

Root caused a production outage

Traced one the team had been chasing the wrong way, then made that whole failure class unrepeatable.

Who did the work
28
Years in engineering leadership
8
Years inside Google Developer Productivity
1
CTO exit, ERP Maestro

Most people selling AI process have never carried the consequences of a slow engineering organization at scale, or operated where "move faster" was a board mandate rather than a slogan. I have, on both sides of the table.

JPMorgan Chase regulated, high stakes, legacy heavy engineering

Ship fast and do not break production have to coexist. The same tension your org lives in daily, without a green field to escape to.

ERP Maestro CTO, exited

Not an advisor. The executive accountable for the technology, through a real exit, and judged on the outcome.

Blizzard consumer scale engineering

Shipping under a quality bar set by millions of concurrent players who notice the moment something ships sloppy. Senior judgment as a shipping requirement, not a phrase.

Google, Developer Productivity eight years

The discipline built around this problem, at the scale of the largest engineering org on earth: how thousands of engineers ship faster without shipping worse.

In their words

JENGAI has real customers and a $2M round closing. The product works. What we needed was to level up how our team builds with AI without falling into the slop trap most teams hit at our stage. Min showed up and did exactly that.

Senior engineering judgment combined with actual AI fluency is the rarest hire of 2026. Mine is named Min.

Shayne Paterson, CEO, JENGAI

Your team is already paying for the tools. Let's get the multiplier.

Tell me what you are shipping and what is getting in the way. The more specific, the more useful I can be on the first call.

Tell me what you're shipping →