Taught by an operator, not a vendor.

Building and shipping software

This is what Unit 3 is taught from. I am not a software developer. What I do is specify, direct, test and ship things I could not have written alone and could not have afforded to commission at a price that made sense for a business this size.

• Eighteen client trip sites, built as static web pages with their own photography and print styling, and deployed.
• A document-intake system. Guests upload passports, visas and signed waivers through the trip site and the files land in the right folder of my drive, through a small server-side function running against a cloud service account. I specified it, directed the build, wrote the setup runbook, and operate it.
• A financial back office. Receipts and payment confirmations read by OCR, renamed to one convention, filed by month and reconciled against bank statements into master workbooks — 457 receipts across seventeen consecutive months — plus chat exports parsed into consolidated tour records.
• Spreadsheet backends, scheduled scrapers and financial tools written as Google Apps Script, which is where the travel company's day-to-day numbers actually live.
• This site. Its pages are generated by a small set of Python modules from one structured curriculum file, and the Schedule Builder on the Tools page was written from scratch for it and tested against fifteen input cases.
• A Ladakhi dictionary app, built from a database recovered out of a discontinued Android app — and I wrote to the publishers asking permission before going any further, offering to delete the lot if they would rather. The technical half of that was the easy half.

Documents, at volume and with consequences

This is what Unit 2 is taught from. Most of it is unglamorous and all of it has to be right.

• Indian foreigner registration. Form C arrival and departure registrations and police checkpoint forms, generated by reading passports, visas, e-visas, entry stamps, tickets and insurance certificates out of photographs, scans and PDFs, and filling the official templates from what they contain.
• Client documents: itineraries, cost sheets, taxi fare estimates, guide checklists and packing lists, produced as web pages and as printable files.
• A full day of STEM teaching for a village school in Zangla, for eighteen or nineteen children between four and twelve: a leader guide, a worksheet, an introduction script and a language-pairing plan for a team in which only two of the adults spoke the children's first language. The rockets were not the hard part.

Research and synthesis

This is what Unit 1 is taught from — a grounded corpus and an open reasoning tool used together, which is the division of labour the unit spends six sessions on.

• Background research for investment and pharmaceutical consultancy work, assembled across NotebookLM, ChatGPT and Perplexity rather than inside any one of them.
• A dormant account traced and recovered. A trust account opened in the 1990s, followed through a fund's merger history to the institution holding it today, with a call script to go with it: the questions in the right order, and what to do when the first answer is no. The method is teachable. The file carries account identifiers and stays private.
• Audio explainers built with NotebookLM, on Indian history and on the Indian healthcare system, made from assembled sources to learn from and then shared. One is public: an episode on Humayun's Tomb, the Mughal garden tomb in Delhi. It is the quickest way to judge this kind of output, because you can listen to it instead of reading my description of it.
• Medical records made legible and portable. A set of hospital records held outside any institution: an MRI study converted from its raw diagnostic format into sixteen labelled sequence images that could be viewed and carried without the hospital's own software, and 127 documents — case sheets, ward notes, lab reports, discharge summaries, prescriptions, much of it handwritten — read by OCR into one searchable, dated record, under a convention that flagged every uncertain reading and every illegible passage rather than guessing at it. The transcript states on its own first page that it is a transcription aid and not a clinical document. That boundary is the entire point of the example: the tools made an existing record complete and portable so the people qualified to make decisions had it in front of them. They made no clinical judgement and were not asked to. Nothing from that work appears on this site in any form.

Design and brand production

• The mark at the top of this page is my own drawing, produced out into a full system: lockups in their light, dark and single-colour forms, icon and social exports, a stylesheet of design tokens and a written set of rules for using them.
• Logos and store graphics for the rental gear business, made with Canva alongside an image model.
• A written brief for putting the same logo problem to other AI tools — a long version for reasoning models, a short one for image generators, and the criteria for judging what came back. The difference between prompting and briefing is most of what Unit 1 teaches, and that document is what it looks like on a real problem.

Further afield

Two things that are not course outcomes and are not offered as any. They are here because the honest picture includes them, and because knowing where these tools stop being useful comes from having pushed them past it.

• Device repair. Diagnostics and firmware reloads on phones and other hardware, run with a model connected to the device through a laptop.
• Training and nutrition tracking against mountaineering goals — the same personal-systems method Unit 2 teaches, applied to something that is nobody's business but mine.