Workman AI Labs · the build log of a one-person AI shop
Builds that survive contact with real use.
Glen Workman builds and runs AI systems for real businesses. We publish tested models, hardware recipes and practical lessons, plus selected business results. Proprietary implementations stay private.
The ledger — dated results, with their evidence and limits explained
Launch architecture Static the corrected week-one log ↗ Editorial claim flags 3 the flag system ↗ Build logs 14 14 published · 0 in draft ↗ Spark-pair decode, ~243K context 63.2 tok/s the spark-pair build log ↗From the build log
All posts → Oct 29, 2026 macfleet-dash v0.10.0: AI engines join the fleet view The October 20 release brings optional SparkDash engine cards, token windows and resource history into macfleet-dash. What the released code adds, and how to read it. Oct 29, 2026 What I would test with a third 128 GB DGX Spark A practical research plan: keep an existing Spark pair available, isolate experiments on a third system, and compare a three-node cluster with a pair plus a worker. Oct 17, 2026 Two AI agents can undo each other’s fixes. The race was in a file write. A recovery helper replaced a genuine resolution record. What the incident and a temporary-state race test taught me about check-then-write, exclusive file creation and the limits of a passing regression. Oct 17, 2026 My AI setup: what runs locally, what goes to the cloud, and what actually does the work Hermes is the working environment, not the model. How I separate a two-DGX local route, cloud inference, bounded tools and a fleet dashboard—and how an actual article moves through that setup. Oct 17, 2026 Your dashboard says a model is running. Is it really? A saved role, an old benchmark and a fresh endpoint report answer different questions. What macfleet-dash v0.9.4 checks, what it leaves unverified, and how I read the result without overstating it. Oct 15, 2026 What daily Hermes use actually looks like: the updater that interrupted the work An October 13–14 operator log, updated October 17: interrupted jobs, unsafe green tests, and a wider claimed/running execution barrier—with separate regression, installed-probe and natural-skip evidence. Oct 15, 2026 My Mac fleet dashboard learned Windows. The hard part was knowing what not to count. What the October 14 macfleet-dash releases actually added: private Windows telemetry, honest pilot states, per-drive warnings, and a clear boundary between released code and unattended monitoring. Oct 13, 2026 Content at scale: what the publication numbers actually prove A content publishing case study: dated evidence of scale, an unresolved inventory mismatch, and why neither volume nor a passing audit establishes business return. Oct 13, 2026 Search visibility: separating accepted URLs from indexed pages A search reporting case study: 1,839 indexed URLs and 4,879 IndexNow acceptances in dated assessments, without treating different cohorts as a conversion funnel or proof of sales. Oct 13, 2026 Deals email: why an attempted send is not a delivery result A Rapid Fire Deals case study in reporting a failed send honestly, keeping activity separate from delivery, and refusing to turn incomplete evidence into a success rate. Oct 13, 2026 Retiring a model the right way: the Qwen decommission checklist Six cron jobs, four dead providers, a dispatcher path, and a pile of presets — the ordered checklist for removing a model from a live fleet with checked client paths and an explicit host-verification gap. Oct 13, 2026 The Spark pair: serving one local model on two NVIDIA DGX Sparks One local-model route across two desk-side DGX Sparks: dated serving measurements, memory failures and repairs, and the limits of GPU-only telemetry. Oct 12, 2026 No invented numbers: the flag system behind this site How this blog decides what gets published: three flags for unproven claims, a status field that does not certify truth, and a house rule that every number carries a receipt. Oct 12, 2026 Week one in public: shipping a static site on Cloudflare Workers The actual build log for this site: an Astro scaffold, a static-assets Worker, one deploy command, the custom-domain cutover — and the receipts for each step.
What the lab runs
Explore the projects & experiments →
Real systems used in our own work. Public case studies share selected results and lessons, not our proprietary architecture, prompts or operating playbooks. Private tools remain private.
- macfleet-dash open source · v0.10.0
- Our open-source fleet dashboard. The v0.10.0 release adds optional AI engine cards, token windows and resource history.
- DGX Spark model lab local AI research
- Our two-Spark test history, plus the proposed research plan for a third 128 GB system.
- The blog-factory public case study
- Our private content-publishing operation. The public case study covers dated results, limits and lessons—not the implementation.
- Rapid Fire Deals public case study
- Our deals publication. The public case study explains why preparation alone does not prove delivery; business rules stay private.
- Search visibility public case study
- A dated assessment of search-indexing and submission results, with clear limits on what those counts establish.
- sparkDash private lane
- Fleet dashboard for the local model lanes — live GPU, memory, and decode metrics.
- Command Center private lane
- The operations cockpit where every lane, project, and worker is run from one place.
I run AI like a shop floor: small, measured, and honest about what broke. We keep the supporting records. Public case studies explain what they prove—and what they do not.