← Engineering index

01 / Atlas · Creator and sole developer

Would you scrap this assembly?

A sealing machine ran at an incorrect setting. Before the shift ended, someone had to decide if the finished bag was still acceptable. I built the agent that brought an engineer and Quality the evidence for that decision.

ATLAS / CASE 01Clinton Williams

The original decision.

Before
An out-of-setting seal made scrap the default.
One working session
The agent assembled tested precedent and the applicable quality requirements.
Disposition
Quality accepted the finished assembly.
Open the case-record extract

The job was to establish whether a past test applied.

A setting tells the reviewer what the machine did. The part’s condition depends on the film, the joint and the range covered by earlier experiments. Atlas searched the seal studies and validation reports, checked the product requirements they addressed and drafted an assessment with citations. Quality checked the match and accepted this assembly in one session. That is one documented outcome. Its unit cost was not recorded.

I built the retrieval, the agent workflow and its evaluation, then demonstrated and marketed it internally. In July 2026 the system logged 1,689 model calls for a 15-engineer alpha group in 18 days, including 442 document-matching calls. One question can trigger several calls. For rollout I care about minutes returned after a person has checked the answer, and which decisions changed.

Disposition / one working session

What could change Quality’s call?

A wrong machine entry establishes a deviation. The question for the reviewer was what it meant for this particular finished part.

Which physical joint was at risk?

The deviation identified a seal made using an incorrect setting. Repeating the nominal specification would tell the reviewer what should have happened. It would leave the condition that actually occurred unexplained.

Had anyone already tested this condition?

Atlas recovered the seal experiments and validation reports and assembled the applicable test procedures and specifications. The review could then ask whether the tested configuration covered this assembly and which quality attributes the excursion could affect.

The use case included particulate and endotoxin requirements among the records checked.

Who could accept the evidence?

Atlas drafted the assessment with citations. Quality reviewed the records, accepted this assembly and avoided scrapping it. A missing configuration match would leave the reviewer with an open question. The account records one bag and gives no unit cost.

Design precedent / a second use case

When does 2.0 mL become a requirement?

The best drainage result on a 5 L prototype was 2.0 mL with deionized water. When Atlas assembled the next design specification, the distinction between a measurement and a promise had to survive.

Original public drawing of the hold-up assessment showing the dry weighing, fill, drain and residual weighing steps.
My hold-up measurement exhibit. Drain conditions and retained mass define what is being compared.
Observed2.0 mL

Best recorded 5 L prototype result with water.

Draft family limit≤2 mL

A proposed retained-liquid requirement across sizes.

5 L stretch target≤1 mL

A separate proposed target. Each intended configuration needs verification.

Source · Atlas-assembled draft design specification, drainage requirement and rationale. Inspect the fields in the draft specification.

Architecture / my design

Why I used two ways to find the same record.

A part number and a description of a failure can refer to the same event. I needed Atlas to handle the engineer who knows the identifier and the engineer who only knows what went wrong.

I combined exact-term and semantic retrieval, merged the rankings and reranked candidate records. A versioned playbook tells the agent what evidence each document type requires. The engineer can inspect the sources that survived the search.

What I replay before changing a model

I run held-out questions against explicit pass/fail criteria. The expensive failure is a plausible assessment that omits the one record showing a different configuration. I would gate a cheaper model on that failure as well as citation accuracy, latency and cost.

Where the reviewer enters

Atlas plans the work, collects evidence, drafts and checks its reasoning. The engineer or Quality owner can reject the conclusion or require new evidence. Demonstrating that boundary was part of getting the system adopted.

A frame from my internal talk, from planning and evidence gathering to a decision by the accountable user.
Original Atlas presentation, 12:40. The person accountable for the part makes the final call.
Clinton’s original town-hall slide shows the agent planning, finding evidence, analyzing, checking itself and drafting before the human decision.
Original town-hall presentation · the workflow I used to introduce Atlas to engineers. The final decision stays with the reviewer.

Operating decision / one month at the same adoption

How many reviewed minutes have to come back?

My July model priced the same normal workload for 133 engineers. The figures below are estimated inference costs at public prices.

ConfigurationCost / engineer / monthReviewed time needed*
As builtabout $45about 22 minutes
Cache repeated contextabout $26about 13 minutes
Cache and route model callsabout $18about 9 minutes
Also halve retrieved contextabout $10about 5 minutes
The break-even arithmeticThe as-built model is $5,948 per month ÷ 133 engineers ≈ $44.72 each. At an assumed $120 per engineer-hour, one minute is $2. The inference cost is covered at about 22.4 net minutes per engineer per month.

My first move would be to cache repeated material and verify the model routing on the same logged questions. Before shortening what the agent reads, I would replay cases with a disqualifying record. If the cheaper version misses it, I keep the longer context for that task. Separately, I would time the human review against the usual workflow.

*Value of available engineering time. Support, infrastructure and change management add costs. Actual invoices, cash savings and net minutes returned were not measured here.  View the telemetry and model inputs