professionalpersonal
500+engineering change orders processed, and counting
An Engineering Change Request (ECR) proposes a change. An Engineering Change Order (ECO) is the controlled instruction that enacts it. Both run on Microsoft 365, with Power Apps for the forms and Power Automate for the routing, carrying a change from design through Quality, Engineering, Production and Regulatory approvals onto the manufacturing floor.
35+new product developments delivered
Product lifecycle management assembled in place of a ~$100K/yr commercial system: SharePoint holds the controlled regulatory documentation, Zoho Projects carries tasks and Gantt scheduling, and Power Automate keeps the two in step over REST and HTTP.
150+controlled documents approved to date
Two separate systems. The FAI agent takes a SOLIDWORKS drawing, extracts every measurement and tolerance on it and fills out our Excel first article inspection matrix — measurement at the quality gate rather than transcription. The documentation agent is the other one: it fills out the DHF (design history file) documents a new product development generates, drafted against ISO 13485.
The count above is the documentation agent's. ISO 13485 is the quality management standard for medical devices, and its document control clause is strict about one thing: a document is not usable until it has been reviewed and approved for adequacy before issue. So the figure counts approvals, not drafts — anything can produce a draft, and only the approved ones reach the floor.
8,000+labels printed since launch
Product inventory and serial number allocation with collision management, a label editor for layout design, and printer proxy configuration driving industrial SATO hardware with a full print audit trail. Runs on a Postgres 17 backend and replaced TEKLYNX, covering GS1 and UDI compliance.
Two brands taken from first brief to launch. Fixed scope, bilingual stakeholders, live products.
25,000synthetic Quebecers · 56 attributes each
A synthetic population of Québec, assembled over 21 build stages from Statistics Canada and Institut de la statistique du Québec distributions. Each stage conditions on the ones before it, so the correlations hold: region constrains language, language constrains likely origin, origin and birth decade constrain the name. From a fixed seed it regenerates identically.
The population then serves as a survey sample — a question goes to a random draw, each persona answers in character through an LLM, and the answers aggregate back into a distribution at a known cost per run. The linked page holds the whole thing up against the census — gaps included — and shows what it answers.
interactive deep dive Explore the population model census validation · live persona draw · cost model · EN/FRAn excavation contractor has a hole in the ground downtown and four places to put what comes out of it. Each site quotes on a different basis — $24 a tonne at one, $200 a load at the next — and each is a different drive across Montreal, at an hourly rate that runs whether the truck is moving or not. MoveDirt answers the only question that pays: which site, leaving at which hour.
$371the cheapest trip at 7h00 — $498 below the next-best site, and a 2.8× spread against the worst. Same dirt, same hour; the gap repeats on every load.
Travel time is queried from the Maps Directions API per departure hour, against Google's historical traffic for that hour in Montreal — so one site prices differently at 7h than at 16h. On an island with the Turcot, the Métropolitaine and a construction season that never ends, the nearest dump is routinely not the cheapest one.
$200 a load reads worse than $24 a tonne until you multiply by a 17-tonne box and get $408. The sites quote in units that don't compare — which is exactly the arithmetic a dispatcher is doing by eye at 6am.
The output is a destination and a departure hour, with the runner-up priced beside it so the call can be overridden rather than obeyed. Exports to CSV, because the office still runs on spreadsheets.
Invite-only pilot: move-dirt.vercel.app redirects to a sign-in wall, so this is a case view rather than a link. Both frames are the running app on a Montreal test project.
