THE BLACKBOARD

How to build a multi-step ReAct agent over personal data: progressive disclosure across three cognitive axes, robust tool call normalization for unreliable LLMs, and a prefetch engine that activates on first interaction.

The Threads API uses a two-step container-then-publish flow, has a critical threading bug where replies must always target the root post, and enforces undocumented content moderation on container creation. Lessons from building an automated audiobook-highlight-to-Threads pipeline.

Migrating from local Ollama to Gemini embedding API cut cost and latency while dramatically improving semantic quality. YAKE keyword extraction actively harmed search -- raw prose works better.

Index of 46 technical books processed by the Skill Forge pipeline into Claude Code skills. Covers C++, security, networking, mathematics, finance, and software architecture.

How to run 10 audit agents in parallel without merge conflicts, and why every async callback in your JS needs a generation counter to avoid stale state.

Two audio pipeline gotchas: Safari kills async audio.play() calls, and FFmpeg -c copy into the wrong container silently crashes. Both cost 30+ minutes to diagnose.

Running vector similarity queries across two PostgreSQL databases from one FastAPI app using asyncpg connection pools, with gotchas around pgvector string serialization.

A voice-triggered pipeline that extracts and transcribes whatever audiobook passage is currently playing. Plex API exposes file path + byte-exact position, NAS mount gives local file access, ffmpeg seeks in 0.15s, mlx_whisper transcribes 2 minutes of audio in 4s. Total latency: under 5 seconds on M2 Pro.

We analyzed commit timestamps across 15+ projects to discover work patterns, then built a tmux config that actually matches how we code.

Globe.gl destroys container children on mount. Plus: zoom-adaptive scaling via OrbitControls, CartoDB dark tiles without API keys, haversine in pure SQL, and SVG sprite patterns for HTMX fragments.

PCI-level rigor applied to personal telemetry. Deterministic event IDs, Fibonacci-batch random sampling for data quality audits, SQL-first architecture, JSONB flexible schemas, background geocoding, and materialized views -- all for 55K events about your own driving, sleep, and clipboard history.

What happens when you apply financial compliance rigor — Fibonacci sampling, API audit logging, SQL-first architecture — to a personal telemetry pipeline.

Cherri compiles text files into signed iOS Shortcuts. The question isn't whether it works -- it's when the investment pays off. Here's the decision framework.

The Cherri compiler turns plain text into signed iOS Shortcuts. Here's what we learned building 8 of them, including a crash bug Apple should fix.

How to make a browser div feel like a Braun T1000 World Receiver. Layered shadows for molded plastic, radial-gradient speaker grilles, CSS seam lines, and why Dieter Rams' 10 principles are better web design rules than most web design rules.

Synology's Docker auto-start API silently fails under memory pressure. The fix is one line in your compose file. Also: NZB obfuscation, container exit codes, and how 23 containers on 4 GB of RAM creates a swap death spiral.

Replaced every emoji across 31 files with Phosphor duotone SVG components. The trick: components for nav/headers, Unicode dingbats for inline text strings.

Ran a full WCAG 2.2 AA audit across every page of bythewei.dev. Found 40+ failures: contrast ratios as low as 1.3:1, missing skip links, no focus indicators, clickable divs unreachable by keyboard. Fixed all of them.

IIBA's BABOK v3 distilled — six knowledge areas, 50 techniques, and the full BA toolkit for stakeholder management, requirements elicitation, and solution evaluation.

Ferrari & Russo's Definitive Guide distilled — evaluation contexts, CALCULATE mechanics, iterators, time intelligence, and the performance debugging mindset for expert DAX.

Evans' DDD distilled — Claude now knows Aggregates from Value Objects, Bounded Contexts from bounded thinking, and can flag Anemic Domain Models on sight.

Bodie/Kane/Marcus distilled — Claude can now price options, build efficient frontiers, apply CAPM, and evaluate portfolio performance with actual math.

Mistral's multimodal audio model is here. the speech AI space just got a lot more interesting.

MCP goes mainstream, agents everywhere, open-source closing the gap