What is underneath: code, numbers and failures, written for whoever has to build it.
Fourteen long notes in two series. Each carries the schema, the query or the formula that was used, the numbers with their assumptions, and the failure-modes section the home page leaves out.
Applied AI · the engineering behind the six
06The home page says what a company can ask for. These notes say what runs, what it is made of, what it costs and how it is measured — with the code.
N.01Scoring where a model belongs: a use-case ledger with a cost per case and a risk class
A department walk-through becomes a ranked ledger of AI use cases, each with a measured baseline, a cost per run and per error, a data-readiness level and an EU AI Act class.
What a non-technical team actually has to understand about a language model
The mechanical facts a non-technical team needs about a language model, plus four checks, four data classes, a versioned prompt library and a benchmark that shows whether it took.
Hybrid retrieval that returns the paragraph, not a paraphrase
Structure-aware chunks, BM25 plus pgvector fused by reciprocal rank, a cross-encoder reranker, and an answer model that must cite the passage id or abstain, all measured on a versioned question set.
An agent can only do what its tools allow, so design the tools
Agents as a tool-surface problem: typed MCP tools with declared effects, idempotent orchestration with explicit state, human approval gates, an append-only audit log and tests that include the refusals.
Serving open models inside the building: sizing, quantisation and the API shape
Weight and KV-cache arithmetic decide what fits on a card, quantisation buys concurrency, vLLM or Ollama by the queue, and an OpenAI-compatible gateway keeps applications portable.
The right size of model for the task, decided by a number
The smallest model that clears a task-set bar on a gold set: candidate ladder, cost per thousand runs, LoRA distillation and a calibrated cascade router.
Systems · working notes
08Notes written from the independent systems and the professional practice: measurement, constraints, and where a number stops meaning what it seems to.
N.07Reconstructing behavioural strategies from public blockchain events
From EVM event logs to positions, windowed features and clusters with a confidence, and the permutation test that decides whether a label is a strategy or noise.
What “under 100 ms” actually means in an event-execution system
A latency claim only means something once the clock's start and end, the percentile and the transaction's fate afterwards are pinned down and measured separately.
Building controlled Linux environments for specialised products
A product image is an enumerated set of packages and services built from a definition, booted read-only, updated atomically with rollback, and provable byte for byte.
Legal constraints as engineering inputs
Under MDR, IEC 62304 and ISO 14971, requirements with ids, risk controls, SOUP records and change control become data the build queries, and CI emits the evidence.
Designing retrieval systems where the source remains authoritative
Citations as resolvable ids, structured extraction with confidence and provenance, an abstention rule and an engineered review queue: the contract between a retrieval system and its reader.
Why a semantic match is not legal eligibility
A similarity score says a call talks about what a company does; eligibility is a conjunction of hard rules evaluated with evidence, and the retriever only proposes candidates.
Meal planning as a constrained optimisation problem
A week of meals modelled as one integer programme in CP-SAT: variables, hard constraints, weighted objectives, a worked instance, and an explanation record built from the same model.
Why LLMs should not calculate nutritional values
A model's nutrient figure is a likely token, not a measurement; composition tables, weights and yields feed a deterministic layer, and the model may quote only what it returned.