Track 01

AI & Technology

Notes from building with modern language models and automation — what held up in practice, what quietly fell over, and what we would do differently.

Focus areas

The ground this track covers.

Applied LLM work

Prompt design, tool use, retrieval, and evaluation for real workloads — the engineering around the model rather than the model itself.

Automation & integration

Connecting services into workflows that run unattended: scheduling, API integration, error handling, and knowing when a human still belongs in the loop.

Data & visualization

Turning messy source material into structures worth querying, and interfaces that let a reader interrogate the data rather than take a summary on trust.

Small, durable software

Static sites, plain files, few dependencies. Systems a single person can still understand and maintain two years later.

Position

A working view of AI

Language models are genuinely useful and routinely oversold, and both facts matter at the same time. The useful part is real: they compress the distance between an idea and a working draft, they read more material than a person has hours for, and they are patient with tedious structure work.

The oversold part is also real. A model will produce a confident answer whether or not it has grounds for one. It does not know what it does not know. Any system built on top of one has to assume that and be designed so a wrong answer is visible and cheap rather than silent and expensive.

The interesting engineering question is rarely which model. It is what you do about the answers you cannot verify.

What that means in practice

  • Ground the output. Retrieval against real sources, with citations the reader can follow, beats an unsourced summary every time.
  • Constrain the surface. A narrow, well-specified task fails in predictable ways. An open-ended one fails in creative ones.
  • Evaluate continuously. If you cannot measure whether a change helped, you are not tuning a system, you are decorating it.
  • Keep a human checkpoint wherever an error would be costly or hard to reverse.

On the tooling

The work behind this site leans on ordinary, boring infrastructure: static HTML, version control, scheduled jobs, and small scripts. Models help produce and check that material; they do not sit in the serving path. A page that renders without JavaScript is a page that still renders in five years.

Coming write-ups

  • Building an analysis dashboard over a fixed corpus — data model, pitfalls, and what the visualization hid.
  • Evaluation harnesses for subjective tasks, and why inter-rater disagreement is a feature.
  • Scheduled agents: the operational reality of software that runs while you sleep.