Five things, built and measured.

problem

I kept losing things I knew I had. I remembered what they were about, never what they were called. Spotlight finds file names, and it misses the video I half watched, the thing I copied last week and the moment someone said it on a call.

constraint

Nothing was allowed to leave the Mac. I built it alone, and it had to run well on my own 8 GB MacBook Air.

decision

1. 01I kept everything on the device: Electron, with native Swift helpers that read text in images, recognise objects and sounds and turn speech into text.
2. 02Search blends keywords with 384 dimensional embeddings through reciprocal rank fusion, then a small quantised ONNX model reranks the top 14.
3. 03It costs $20 once, with a signed offline licence. No account, no subscription, nothing to phone home to.

problem

Parents, schools, early learning centres, organisations and vendors were ordering and preparing meals across separate tools. And the platform holds children's data, so every shortcut has a cost.

constraint

It's live in several countries, each with its own privacy law. There are apps in both stores, and every release has to stay shippable.

decision

1. 01I ran security as engineering work. I led the ISO 27001 implementation, covering the risk assessment, access control and audit evidence, and mapped it to each national privacy law.
2. 02I own the architecture and give the final approval on every feature, across Next.js, Python, Node.js and .NET services.
3. 03I grew it from school meals into early learning centres and organisations. That release is the newest one, and it ships alongside the React Native app on iOS and Android.

problem

The boutique needed a real shop with categories, search, a wishlist and delivery across Mauritius, without sending customers through messages to order.

constraint

Customers in Mauritius pay by Juice or bank transfer rather than card. Delivery has to reach Rodrigues too, and the owner manages the product photos.

decision

1. 01A Next.js storefront, with Supabase holding the catalogue and the product images at responsive sizes.
2. 02Checkout is built around Juice and bank transfer, and delivery is free over Rs 2,500.
3. 03I kept the type quiet and editorial, so the clothes carry the page.

problem

Most small shops around me run on WhatsApp, paper and Excel. Invoices get lost, stock runs out without warning, and nobody remembers who still owes money.

constraint

I built it alone. Shop owners won't buy POS hardware, so it has to run on the phone they already carry, in English, French and Kreol.

decision

1. 01One app covers counter sales, barcode stock, invoices and customers, as a PWA on the phone.
2. 02Invoices go out as a WhatsApp link or a PDF. It takes cash, card and MCB Juice at the counter.
3. 03I launched it free, with no card and no trial clock. It runs on Next.js, Prisma and PostgreSQL, with cookie free analytics.

problem

Ask an assistant model what it is and you get a trained script. I wanted to know what is underneath once the system prompt, the chat template and the instruction tuning are gone.

constraint

I had one day, an 8 GB laptop for the 3B model and one rented A100 80 GB for the 32B. The samples were small, so the paper could not claim more than it shows.

decision

1. 01I used base Qwen2.5 models only, with no system prompt and no chat template.
2. 02I gave the model a life: memory of every event, its own surprise as a signal, self set goals, read only senses and LoRA learning on its own writing.
3. 03I ablated internal directions and recorded every failure alongside the results.