Research & field notes

Experiments run for real. Reported as they landed.

Autonomous agents, local models, backup systems, migrations — the working notes behind the practice. Successes, failures, and the constraints that decided them, published in full.

Oct 2026Testing Timestral on the CAC 40: what a time series model does when the data is noiseI evaluated Mistral AI's Timestral-209M zero-shot forecasting model on the hardest test I could think of: daily predictions for all 40 CAC 40 stocks. The model did not beat a one-line persistence rule, but its uncertainty envelope stayed calibrated, and that calibration is the product. Notes on the sealed cast protocol, the COVID crash regime analysis, and why unpredictable data is the right stress test.Oct 2026Building a harness for a 4B model: the recipe made it repeatable, and the failures got quieterA two-round experiment to turn the llm-wiki skill into an enforceable skill for small models (slm-wiki): Gemma 4 E4B fails a wiki task in pi, a local Qwen 3.8 27B writes an AGENTS.md recipe as the harness, the 4B re-runs with the recipe and produces repeatable results, and the analysis shows the failures got quieter rather than fewer. Closes with gated skills, where the harness enforces the guardrails the model skips.Sep 2026Three local models, one 96-minute session: the 35B corrupted the file, the 27B fixed it, and the 9B distill shipped 70%One session, one task, three local models: a 35B MoE that corrupted a chat template by retyping it, a 27B dense model that fixed it with curl and cp, and a 9B distillation that executed well but never self-corrected. Notes on verification, retyping as a failure mode, and supervision cost.Sep 2026Tool calling with small local models: why the Qwen3.5-9B failed a task built for the LFM2.5-230M (and the `thought` field fix)Field notes from a controlled experiment on local LLM tool calling with small open models: why context format beat model size, why a constrained `thought` field is the highest-leverage change, and where the capability wall actually sits.Sep 2026The New Hires Never SleepA one-human consulting firm puts four AI agents on the payroll: a 412-second reply, a 401 to itself, a notes API that lies politely - and the three rules that survived the first real week.Jun 2026The G.E.A.R. experiment: 47 hours of autonomous AI agent, what broke, and what I learnedBehind the demos of OpenClaw and Hermes, here is what happens when you actually run an autonomous AI agent for 47 hours without stopping.Mar 2026CRM migration with local coding agents: why the M5 Max MacBook Pro mattersField note on using local coding agents for CRM migration and sensitive database workflows, and why the M5 Max MacBook Pro memory profile matters.Mar 2026How to run OpenCode with LM Studio locally (Qwen3.5 on a 6-year-old MacBook M1)Step-by-step field note for running OpenCode with LM Studio locally using Qwen3.5 9B and 0.8B, including Telegram bot workflow, local config, and M1 performance tradeoffs.Feb 2026How to think about backups: gzip.cloud design choices and rationalesDay-zero design notes for a plain, reliable backup service: universal SFTP storage, immutable backups, reminder-based ops, and restore certainty.Feb 2026Ensuring resilience from day one : Building gzip.cloudThe minimal remote backup service I want to run for gzip.cloud, designed for reliability, low ops, and client-side encryption.

The archives

Earlier technical resources, essays, and tutorials — the layer this practice is built on.

Browse the archives

Get the weekly insights — free Who is behind this