Signal-to-system playbooks
Custom Model Training
Practical guidance for teams deciding when a frontier signal deserves a trained model, a dataset, an eval gate, or a safer rollout path.
Signal-to-system playbooks
Practical guidance for teams deciding when a frontier signal deserves a trained model, a dataset, an eval gate, or a safer rollout path.
Start with decision quality, then move into data quality, then enforce ruthless evaluation before launch.
A field note on local vision-model workflows, examples, and why cheap local AI unlocks slower but useful inspection loops.
Read field note →A field note on using local vision segmentation, cheap inspection loops, and agentic review tooling to eyeball an early Independence Town build.
Read field note →How to decide between prompting, RAG, fine-tuning, or full custom training without burning six weeks for a 4% gain.
Read playbook →Build a dataset that reflects real user behavior, edge cases, and failure modes instead of happy-path vanity examples.
Read playbook →A practical eval stack: offline evals, scenario tests, red-team checks, and hard release thresholds that block bad launches.
Read playbook →A plain-language decision matrix for selecting the lightest approach that achieves your target behavior and reliability.
Read playbook →Field notes and a four-part series connecting Eric's local-inference experiments, pico-LLM work, nanochat, small-language-model research, and practical ability training.
A Jetson Nano field note on local AI economics, always-on generation, payback periods, and when edge inference starts saving more than it costs.
Read field note →A prompt-craft field note on queued agents, latent-space gardening, and pruning model work into practical operating loops.
Read field note →A field note on why agent-native Kanban, triage, and human unblock points make multi-step AI work feel less like Jira and more like a live operating system.
Read field note →A project-writing field note on capturing lessons while the system is still warm, drafting ahead, and turning publication into a reusable learning loop.
Read field note →A short model-training note on using the generated artifact itself to expose pixel-art technique cues the model can reuse.
Read field note →A short SLM field note on handing the GPU to an OpenClaw instance for nanochat experiments in intelligence per megabyte.
Read field note →A first look at the Snapdragon Elite X1E80100 Qualcomm Hexagon NPU as a local-inference path for low-cost agentic tokens.
Read field note →A rainy-day local-hardware note on testing whether a Jetson Nano can carry useful learning loops close to the bench.
Read field note →A short local-inference field note on a tiny Qwen model sketching a full C# app plan on older NVIDIA hardware in under a second.
Read field note →A field note on custom Qwen LoRAs, Benjamin Franklin model voice, pseudo-SWE-Bench experiments, and narrow-model usefulness.
Read field note →A short model-training note on constraining weights to -1, 0, or 1 and using overnight run stats to pressure-test generalization.
Read field note →A local-inference field note on Gemma, ONNX, battery-powered AI, and the harness needed to turn a small local model into a useful agent.
Read field note →A debugging field note on 3D pipeline views, runtime charts, and tooling ideas for finding where inference performance actually drops.
Read field note →A QNN field note on HTP, HMX, HVX, ONNX graph partitioning, and why Gemma4 workloads may still route through a hot CPU.
Read field note →A short benchmark note on why selected NPU runs can still show CPU token work when prompt prefill has not moved to the accelerator.
Read field note →A local Snapdragon benchmark note showing how a Gemma vision harness compares CPU, GPU, and NPU profiles on year-old hardware.
Read field note →A Gemma4 field note on moving a custom Snapdragon harness from slow token/sec baselines toward usable GPU-backed local inference.
Read field note →A short field note on Snapdragon NPU experiments, local Gemma sessions, and the slash commands that make a tiny-model harness usable.
Read field note →A short progress note on making local Gemma4 workloads target the GPU instead of drifting into CPU fallback.
Read field note →A local Gemma4 vision harness for progressive screenshot inspection, ONNX sessions, and cheap edge inference across GPU, NPU, and CPU targets.
Read field note →A benchmark field note on why CPU, GPU, and NPU numbers can tell a surprising story until the ONNX graph, kernels, and DirectML path are made visible.
Read field note →The preserved 64 MB RPG-state experiment reframed around constrained domains, token contracts, failure-first evals, and capability-per-megabyte.
Read part 1 →How nanochat's depth dial turns training into a comparable family of compute-optimal models instead of one-off checkpoint luck.
Read part 2 →What changes when GPT-2-level capability becomes cheap enough to repeat, and why data and evals become the real constraint.
Read part 3 →How synthetic data, token-visible task design, and identity tuning turn small models into useful narrow specialists.
Read part 4 →Source archive · July 2026
Short experiments belong beside the durable playbooks when they show the actual hardware, models, and failure modes behind the work.
A public visual record from a local image-model experiment.
Open field note →A compact record of a constrained-hardware image workflow.
Open field note →An honest training note: a small experiment, a crash, and the evidence retained from both.
Open field note →