The MechAI log
Notes from the seam between physics and machine learning.
Working notes on forecastability, predictive maintenance, digital twins, and shipping AI into industrial environments — plus a few interactive decision wizards below.
- Oct 27, 2026
Your Agent Can't Grade Its Own Homework: What Reflexion Actually Showed
An agent improves only as far as a signal it does not produce lets it. Reflexion's gains rested mainly on the pass/fail check, not the reflection, and plain retry beat it for less money.10 min read#Agents#AIEngineering#Research - Oct 13, 2026
MechAI Is on YouTube
I'm starting a YouTube channel. Tech in general, but mostly aviation: how aircraft systems work, and where AI actually helps keep them flying.1 min read#Aviation#YouTube#Announcement - Sep 18, 2026
The Fix Is the Pattern
A pattern name buys you nothing. The step budget, the observation check, the reject path, and the human interrupt are what actually stop an agent from looping, leaking, or lying. 46 LangGraph traces, with token counts from the traces.16 min read#Agents#AIEngineering#Architecture#LangGraph - Aug 20, 2026
Show Me Your Graph. Tell Me You're a Rookie.
A graph is not an architecture, it is a control flow representation. When an LLM agent framework becomes the default shape, the diagram answers the design question for you.11 min read#Architecture#AIEngineering#Agents - Jul 31, 2026
Choose Evidence, Not a Shape
How to design a testing strategy from risks, evidence, and cost — not pyramids, trophies, or prescribed ratios.17 min read#Testing#Quality#Interactive - Jun 14, 2026
arch-crew: a plugin for deciding (and auditing) architecture
Three Claude Code skills that help you pick — or assess — software architecture, design patterns, and agentic systems, recommending the least architecture that meets the requirement. They grew out of three interactive decision wizards you can click through here.4 min read#Architecture#Claude#Interactive - May 28, 2026
Is It Even Forecastable? Triage Before You Model
Most forecasting projects start with model selection. They should start with a cheaper question — does this series contain exploitable structure at the horizon you care about?2 min read#Forecastability#TimeSeries#OpenSource - May 12, 2026
A Digital Twin Is Not a Fancy Dashboard
In manufacturing, most 'digital twins' are real-time monitoring dressed up with 3D visuals. A real twin answers a different question: what happens before you change anything?2 min read#DigitalTwins#PredictiveMaintenance#Manufacturing - Apr 15, 2026
Remaining Useful Life: The Model Is the Easy Part
RUL estimation for industrial assets looks like a regression problem. The hard parts are the label, the cost of being wrong, and whether an operator will trust the number.2 min read#PredictiveMaintenance#RUL#Reliability - Mar 2, 2026
From Heat Sinks to Backprop: An Engineer's Path into AI
The transition from mechanical engineering to AI isn't a pivot away from physics — it's physics gaining a new instrument.1 min read#Career#Engineering#Reflection
Interactive decision wizards
Click through a decision tree yourself — or, if you're less sure, install the arch-crew skill and let it interview you.
Software Architecture Patterns
Click through the selection tree — structure, topology, the data-platform taxonomy, overlays — to the least architecture that meets your requirement.#architecture#interactiveOpen wizard ↗Agentic System Patterns
Design or audit an LLM-agent system layer by layer: autonomy, reasoning loop, topology, memory, reliability, integration.#agents#interactiveOpen wizard ↗Design Patterns (GoF / Pythonic)
Which-pattern interview to a single recommendation — with a disambiguation step and its idiomatic, modern-Python form.#patterns#interactiveOpen wizard ↗Testing Strategy Builder
Build a risk-driven portfolio of quality practices, automated tests, and evaluations — starting from the failure you actually need to catch, not from a pyramid.#testing#interactiveOpen wizard ↗