Posts

On the engineering judgement behind these projects: why each call went the way it did, what it gave up, and what is still not good enough. Available in English and Chinese.

1. 15 August 2026·About 4 min read

   Consent is not a checkbox: designing for highly sensitive data

   GDPR Art. 7 sets the conditions under which consent counts at all. The real bar is not more explanatory text — it is whether your product still works properly after someone refuses.

   • privacy
   • compliance

2. 15 August 2026·About 4 min read

   Grounding an LLM with RAG: the hard part is deciding when to retrieve

   Retrieval stops a model inventing clinical advice. The harder question is when to retrieve at all — and the two ways of getting that threshold wrong cost very different things.

   • rag
   • llm

3. 15 August 2026·About 5 min read

   Why Random Forest beat XGBoost and an LSTM on my dataset

   Forecasting mood from a behavioural time series points straight at an LSTM. With a few hundred rows per user, a 50 ms latency budget and users who ask why, it points somewhere else entirely.

   • machine-learning
   • model-selection