Your AI prototype works. It still isn't in production.
Six months in, every team says their piece is fine, and nobody can tell you what is actually blocking the ship date.
If you have said one of these out loud
- The demo landed in March. It is still sitting there.
- Every team swears their part works.
- Nobody can tell me what "done" means for this thing.
- We cannot say what happens when it is wrong.
- If the person who built it leaves, we are finished.
What I found building 77 Rules
Every component passed. The system still could not be trusted.
I built 77 Rules in four 60-hour weeks: an AI system that intensively reviews and grades dashboards for best design practices, data integrity and other common errors. It has been implemented onsite with clients.
One model doing everything performed poorly. It missed important issues and created false positives. Through iterative testing, I split the work across models based on what each could do reliably and economically, then used a stronger model to verify final flagged issues. The selected production configuration reduced inference cost by approximately 65% while retaining approximately 98% of the leading model's benchmark score.
The production lesson: model choice is not a single decision. It is an architecture decision.
That failure is invisible from inside any one role. It only shows up when you look across all of them at once.
The mechanism
Production AI fails in the space between five roles.
Not inside the model. Not inside anyone's code. In the handoffs, where one seat assumes another already handled it, or where a clear explanation changes meaning as it crosses roles.
Every seat can be locally right and the system can still be wrong. Nobody owns the space between them. That space is where forward-deployed engineering earns quite a bit of leverage, in my estimate about 2x faster with a higher quality outcome likely.
I have worked across three of these seats and spent decades working closely with all five. Ten years of product management leadership at Brio, SAS, Yahoo, Microsoft, and Tableau. Eight years of clinical-trial work under FDA validation. Designing and building enterprise data warehouses since 1994. Leading large development teams, including a 40-person team at SAS for 5 years.The full record, seat by seat →
The shift
Change the question and the blocker appears.
"Does this component work?"
Answerable by each team, in isolation, honestly, yes. And the project still does not ship.
"Which production capability is still missing, lacking or just hard to use?"
This question makes the blocker, the owner, and the next action clear much faster.
Two ways to work with me.
If you are the one building it
The YakData Production Handoff
Six weeks. Bring the system you are already building at work and learn to examine it from all five production perspectives: product, UX, front end, back end, and QA. You leave with six concrete working documents that expose what is still unresolved, who owns it, and what has to happen before the system is ready to ship.
$2,950Founding cohort · 10 seats · Sep 21 to Oct 28
If you want YakData to own the delivery
A private YakData engagement
Start with a two-week Production Read when you need a clear diagnosis. Use Forward Deployment when you want me to define, architect, and build the first working version with your team. Or keep me involved as Standing Principal while your team ships.
From $9,500Diagnosis · Build · Ongoing principal review
Where the judgment comes from
Systems I built, not slides I made.
Stephen looks at what the business needs and works back to the technology. This insight is invaluable for improving ROI and driving quick wins from the analytical investment.John LodmellCFO
Nordstrom Credit Services
Before you talk to anyone
Put a number on the delay.
Team size, loaded cost, months elapsed. Ninety seconds. Most people have never seen the figure, because it lives in four separate budgets.
Private engagements
From a two-week diagnosis to a working system.
I can diagnose what is blocking production, build the first working version, or stay with your team as the principal-level technical reviewer while they ship. Forward Deployment is the build engagement: I define the decision, architect the system, and build one complete end-to-end version. Your team owns the code and takes it forward.
Production Read
$9,500Two weeks
I read what exists, interview four to six of your people, and give you a written answer to three questions: Is this on a production path? What is blocking it? What should happen next? Includes a 90-minute sponsor readout.
Build engagement
Forward Deployment
$78,000Six weeks · $13K/week
I embed with the team, define the decision, architect the system, and build the first working version: one real transaction running end to end in your environment, with evaluation, failure behavior, and the operating controls needed to trust it. Your team receives the code, tests, architecture, and handoff package and owns it from there.
Standing Principal
$18,000/moThree-month minimum · Two clients maximum
I stay across the production decisions that are expensive to get wrong while your team builds: architecture, model routing, evaluation, failure behavior, release readiness, and handoff. Weekly working session plus review of consequential decisions between sessions before they harden into rework.
Start with a Read. If the right next step is Forward Deployment and we start within 30 days, I credit the full $9,500.
Tell me what needs to work in production.
A few precise sentences is enough. First contact is for fit and scope only.