Archives

April was the month the platform stopped being one fast thing and started being four uneven ones. The async refactor made the fast paths embarrassingly fast, but deep search hit a ceiling — and chasing it meant discovering that the knowledge graph underneath two of my features had been empty for weeks, the autonomous fleet had been starving on it, and the connector pipeline I'd been proud of had a producer shouting into a consumer that was never built.

A Monday that converted the whole RAG pipeline from sync to async (fast search 9 seconds to under 150 milliseconds), then watched the platform's own automation find a path traversal in its own code — bracketed by the discovery that an upstream auto-update bug had quietly deleted two months of the evidence this series is written from.

Every AI coding agent signs its commits with a forgeable plain-text line. I gave each of mine a non-exportable key in the Mac's Secure Enclave, hook-enforced, with a verifier that flags forgery — here's the build.

A weekend search-quality sprint that took benchmark scores from 0.67 to 0.74, an MCP transport migration that cut cold starts by 96%, ten bugs found along the way, and a CVSS 9.4 Harbor CVE caught by an automation pipeline deployed the same day — with the benchmark research that made it all measurable.

A false-alarm audit on Wednesday patched 37 CVEs, reverted a cluster upgrade in fourteen minutes, then cascaded into ten more tickets on Thursday. Twelve tickets, twenty-six hours, one lesson.

How the second knowledge graph went from intermittent to dependable — Python AST plus FalkorDB plus Leiden community detection — and the moment the Kafka consumer pattern from the journal pipeline generalised into a fleet of code-intelligence workers. With one deliberate reversal from LLM-based to deterministic NLP that cut hallucinated tags from 75% to 0.3%.

How a single score threshold became the difference between a bot that hallucinates and a bot you can trust — and why the architectural decision to refuse synthesis is the same shape as a circuit breaker, a feature flag, and a dead-letter queue.

I exposed my MCP bridge to the internet so Claude.ai could search my vault remotely. Within 26 hours, Cloudflare logs showed 39 searches from 15+ Anthropic IPs — and I had no way to tell what they'd asked for. Here's the incident response that sealed every secret, obfuscated every endpoint, and bootstrapped a proper engineering workflow in the process.

The automated journal entries were fiction. Every single one. Here's how a broken pipeline got replaced in 10 days — Kafka, vLLM workers, nine code iterations, a context preservation fix, and an LLM-as-Judge quality gate — all tracked through git commits and vault evidence.

3.5 months of silence hid 450 commits, a platform rebrand, 2TB migration, and the transformation from side project to platform engineering discipline.

BT SRE interview Oct 17th morning (strong networking/Linux, exposed K8s gaps), followed by 5.5 hours implementing Anthropic patterns after work. Achieved PERFECT 100/100 alignment (66→94→100) via parallel execution, enhanced MCP docs, LLM-as-judge, bash tools, subagent pattern, resumable execution. Skills migration Day 5-7 complete (103 tests, 99.4% token reduction). The honest gap - AI-assisted execution masked foundational knowledge. Claude Code rebuilt K3s cluster, but I couldn't explain control plane in interview. Manual rebuild plan to close the gap.

Working with Claude, we shipped three infrastructure systems in 48 hours. Neural Vault crisis forced reindex of 41K docs. Built Journal v3 (8-task pipeline), Hybrid Search (A/B tested), and 7 slash commands - all behind feature flags. Parallel testing revealed hybrid search only 1.7% better despite 27x slowdown. When you can't test everything manually, you architect for failure.

54 minutes of active work across 14 hours (3 sessions with gap-aware timeline). Six journal automation enhancements deployed - RAG context (30-day search), brag docs (61 accomplishments), emotional tracking (burnout detection), interstitial notes (real-time capture), semantic chunking (102 chunks), automated pipeline. ULTRATHINK diagnosed Prefect silent failure in 45 minutes. Deployed hybrid Cron+Prefect fix in 12 minutes. Created idea-tree system for fast visual brainstorming. Built /blog command with 5-gate quality system. Automation begets automation at maximum meta-level.

18.5 hours across 48 hours. Research-backed codebase review (85/100), 5/5 systematic tests, NER enhanced to 163x speedup, 2.5 hours metadata debugging. Not just deployed - validated and improved. The research foundation behind every enhancement. From 85→90 quality score through methodical engineering.

5 hours debugging temporal search to validate Season 1. Found 0 files from Sept 11, then 9. Found 55 files from Oct 5, wrote 7 episodes. Compression ratio discovered. Then got caught fabricating the timeline in this very episode. Meta level maximum.

The final episode: Using ConvoCanvas to document its own creation, then publishing with AstroPaper and seamless episode navigation.

From hand-coded Mermaid diagrams to automated architecture visualization. An evening working with Claude to teach the system to document itself.

Saturday morning, October 5, 2025. A K3s cluster silently failing for days. 6,812 pod restarts. Working with Claude to diagnose and rebuild everything.

K3s was already running. The question wasn't whether to use it - it was whether to migrate everything to it. A day of deep research revealed the hidden cost of containerization.

A weekend implementing ChromaDB, a Monday demoing it at work, and discovering that 1,142 files need more than folders - they need semantic search.

From context window frustration to local control. September 18-19, 2025 - a weekend working with Claude Code to install 17 local models, discovering what it means to own your AI conversations while working a full-time job.

From vault design to working code. September 14-15, 2025 - vault creation, LibreChat deployment, and the first FastAPI backend that could parse conversations and generate content ideas.

The context window error that sparked everything. September 11, 2025 - an evening planning session with Claude that would shape 25 days of building.