Engineering notebook / source extracts
The project records.
These are public, anonymized extracts from my work. Each names the underlying record, shows the part that changes a decision, and states what the record can support. Original public project pages are linked where available.
Atlas / use-case showcase · Case 01
What Quality reviewed.
An anonymized extract of my written production case. The manufacturing record and internal documents identify the assembly and are private.
| Case field | What the record says |
|---|---|
| Situation | A finished bag was sealed using incorrect parameters. Scrap was the default. |
| Work performed | Atlas found the seal experiments and validation reports for the condition. It cross-referenced endotoxin and particulate procedures and specifications and drafted a cited engineering assessment. |
| Outcome | Quality owned the disposition and accepted the bag. The case account records one assembly, one session and no unit cost. |
Source locator · Atlas use-case showcase, Case 01, Situation / Action / Result. Read my design decisions
Atlas / July telemetry and inference forecast
What the counter actually counted.
5–23 July 2026. The source defines an 18-day window and an alpha group of 15 engineers.
| Source field | Recorded or calculated |
|---|---|
| Total recorded model calls | 1,689 across seven models. |
| Embedding calls | 442 calls that turn text into a representation used to find related documents. |
| Other model calls | 1,689 − 442 = 1,247. Several calls can serve one user task. |
| Public-list estimate | $273.55 for the observed mix of calls. Contract invoices and verified labor savings were outside this calculation. |
The July forecast’s normal-adoption row assumes 133 engineers. It estimates $5,948 per month as built and $1,347 after all modeled changes. At the same assumed labor rate of $120/hour, the inference-only thresholds are 22.4 and 5.1 net minutes per engineer per month. A cheaper configuration must still recover the evidence a reviewer needs.
Source locator · Atlas Inference Cost Forecast, July 23, 2026, sections 01, 03 and 07. Figures are drawn from telemetry and a separate same-adoption price model. Open the comparison
Atlas / draft design specification · DRN-001
One measured result, two proposed limits.
A drafted requirement can preserve the best earlier result and still remain unverified.
| Field | Value and scope |
|---|---|
| Precedent | Best reported first-generation 5 L prototype result: 2.0 mL with deionized water. |
| Draft requirement | ≤2 mL retained liquid across the proposed family. |
| 5 L target | ≤1 mL for the 5 L configuration only. |
| Verification state | Not yet verified. A gravity-drain test method was to be written. |
Source locator · Atlas-assembled second-generation final-fill header bag specification, drainage row DRN-001 and rationale. See the method illustration
Film-to-film bonding / Rev F, original report
Which specimens passed?
The measurement, functional result and revision history come from three different parts of the report.
| Record | Observed | Decision boundary |
|---|---|---|
| Figure 17 · one cycle | Four channel peaks: 244.9, 244.2, 228.9 and 229.7 °F. Hottest − coldest = 16.0 °F. | A single cycle does not establish a safe window or a fixed hot location. |
| Table 14 · September 24–25, 2026 | 1 L: 18/18 passed. 50 L: 15/15 passed. Stepped hold to about 6.5 psi. | The forms did not record the bond settings or dwell at each pressure step. |
| Appendix D · Rev D → Rev F | The later 10-run screen recorded T2 hottest in 7/10 runs. The earlier fixed-hot-spot and uniformity-only interpretations were withdrawn. | Recipe and setup variation coexist. A cause or release boundary needs linked functional results. |
Source locator · Film-to-film DOE, Rev F, Figure 17, Table 14 and Appendix D, Table 27. The summary omits equipment identifiers and production recipes. Open the original trace
Customer validation / September 2026 project account
72 samples could have qualified the outgoing port.
Two programs were drawing on the same part at different times. The product team was about to build customer validation samples while Manufacturing was qualifying the new molded port for launch.
I spotted the conflict before the sample build and negotiated parts from the new validation. All 72 customer samples shipped on time with the correct revision. Rebuilding, reshipping and asking the customer to repeat their work was estimated to expose a month or more.
Source locator · September 2026 engineering role review, freeze-thaw customer builds / molded port. The schedule exposure is a counterfactual. The shipment is reported. Open the project
CAD assembly / component survey
37 usable parts out of which denominator?
The software could inspect 39 occurrences in an assembly of 199. The other 160 were loaded in a lightweight state and their metadata were unknown.
Two inspected occurrences lacked usable metadata. The correct next run fully loads the assembly and deduplicates reused parts. An assumed two to four hours of preparation per missing part cannot be multiplied by all occurrences as if each were a distinct purchased part.
Source locator · CAD assembly agent proposal, mixer component inventory and feasibility assessment. Open the engineering record
69 runs left two explanations.
Stroke and temperature changed together in my membrane-to-polycarbonate screening study. Increased closure can improve physical contact and heat transfer. More heat can also improve the joint. Repeating the same coupled settings would make their combined effect more precise while leaving the cause unsettled.
My next experiment would hold temperature constant while varying closure, then reverse the comparison. I would record the interface temperature and where the tested bond fails.
Source locator · 69-run screening account and subsequent bond-window model review. Open the project
The large pack had about $5 left.
The incumbent protective box cost about $60. Material sums for a three-module prototype were $54.33 and $54.81 on two preliminary estimates at a 5,000-piece component quote basis. That left roughly $5 before assembly labor, extra tubing restraint, freight and qualification.
At an illustrative loaded labor rate of $60/hour, five extra minutes consumes $5. The 60 lb transit assembly survived, but the tubing expanded past the divider. I would fund a restraint redesign and a repeat of that packout before authorizing tooling. The product did not launch.
Sources · August 2018 concept review, slides 5, 7 and 11–12. Quote sums are estimated materials. The comparator is a purchased box price. Original transit account and photographs
Can the camera see the clamp?
A clear clamp on reflective steel can lose its visible edge. The left image is from the held-out set. The right is its hand-labeled reference outline. That pair shows what the system should find. The inventory of 74 matched image-label pairs gives no production escape rate.
Source locator · Automated assembly inspection, 2024 held-out image and corresponding reference. Open the inspection record
Genysys / build history
Software checks and a live athlete are different evidence.
The verified history supports a working founder-pilot foundation. It does not supply a completed live episode or a paying customer.
| Source | What it establishes |
|---|---|
| July 25, 2026 · merged founder-pilot change | Wearable sync and canonical observations, Android plan entry, separate commitment and execution, later outcome review and preserved evidence. The engineering record reports 174 backend tests and 67/67 release evaluations. |
| October 31 · open build contract | A frozen evidence record, deterministic reconciliation and historical replay are specified. The implementation issue remained open at the November review. |
| Original product walkthrough | The interface depicts a recommendation, the athlete’s choice and the next session. The depicted interaction is illustrative. |
Source locator · Genysys founder-pilot merge #10, event-compiler issue #12 and original Arciv product walkthrough. The source repository is private. Read the proposed pilot