The Route Had to Reach the Model Season 5, Episode 3
Engineering in the open
Ryan Duffy
Site reliability engineer who builds AI infrastructure. A global network by day, a self-hosted AI platform by night — and the decisions written down along the way.
Latest from the journal
The Route Had to Reach the Model
More local context meant little until requests reached a working path. Following the router, its override, an empty-result cache and the evidence used to tune them.
Read the episodeLatest updates
1 website commit Website · WK·45
1 commit — Merge pull request #35: Publish Season 5 Episode 3
The Smaller Model Had to Earn Its Place Season 5, Episode 2
The Job Was Green. The Graph Was Not. Season 5, Episode 1
The Self-Healing Turn Season 4, Episode 19
An agent investigates. An SRE decides. ARTICLE · 12 minutes
The Ops Day Season 4, Episode 18
Wiring the Trust Engine into This Blog Season 4, Episode 17
The Trust Engine — the Grounding Architecture Season 4, Episode 16
Governance-as-Code — the Gate That Gates Itself Season 4, Episode 15
Measuring the Pipeline — CI Fires + DORA Season 4, Episode 14
The Second Restructure — Cutover Waves + the Hero Bug Season 4, Episode 12 · 16 minutes
Fix One Layer, Expose the Next Season 4, Episode 13
The Dreaming Loop Season 4, Episode 11 · 17 minutes
Feeding from Outside Season 4, Episode 10 · 14 minutes
The Night the Mac Died Season 4, Episode 9 · 20 minutes
The Number It Was Reading Season 4, Episode 8 · 13 minutes
The Proxy That Ran Hot Season 4, Episode 7 · 13 minutes
The Codex Lane Season 4, Episode 6 · 16 minutes
I Won the Writing Award. Quên Won the Track. What I Learned About Production AI ARTICLE · 12 minutes
The Instrument Panel — Voice, Runner v2, Terminal Season 4, Episode 5 · 13 minutes
Building the Cockpit — Thursday Chaos to Console-Born Season 4, Episode 4 · 12 minutes
Git History for the Agent's Brain Season 4, Episode 3 · 19 minutes
The Morning After — Proving the Fixes Were Real Season 4, Episode 2 · 13 minutes
Memory as a Measurable Engineering Problem ARTICLE · 8 minutes
The Night the Fleet Audited Everything Season 4, Episode 1 · 13 minutes
The Great Restructure — rebuilding the skeleton while the platform learns to measure itself Season 3, Episode 10 · 19 minutes
I Rebuilt My Site and Opened My Decision Log ARTICLE · 9 minutes
Self-Healing SRE — LLM Autonomy Ladder ADR-088 · Accepted
LLM Content Trust Boundary — Four-Layer Defence with End-to-End Provenance ADR-047 · Accepted
Severity-Routed Discord Alerting via Self-Built masgustavos Forwarder ADR-087 · Accepted
GPU/WebGL Graph Lens over FalkorDB (rootweaver-console Graph Tab) ADR-086 · Accepted
The Speedometer I Made Up: Building a Token Budget Governor for AI Coding Agents ARTICLE · 16 minutes
Scout Fleet and the Async Ceiling — what a 200× speedup doesn't fix Season 3, Episode 9 · 18 minutes
Sever the neural-vault Import Blind Spot (vs Workspace-ify It) ADR-083 · Accepted
Vault Ingest Lanes and Lane-Aware Retrieval ADR-082 · Accepted
Cluster-Wide Namespace Taxonomy and rootweaver Decomposition ADR-080 · Accepted
Evaluation Harness Governance Framework ADR-081 · Accepted
The defenses you haven't built yet — async, a path traversal, and the bug erasing the evidence Season 3, Episode 8 · 16 minutes
Inter-Hop Context Compaction for Multi-Hop Reasoning ADR-077 · Accepted
Claude-drives-Codex Autonomous Handoff (Dual-Agent Coding Routing) ADR-079 · Accepted
Workload-Level Provenance for Autonomous Agent Commits ADR-075 · Proposed
Agent-Ops Console Module — Desktop Session Runner with Template-Gated Launch ADR-072 · Accepted
Interactive Agent Sessions — Audited Free-Form Supersedes Template-Only ADR-073 · Accepted
AI Agent Subscription Budget Governance (Claude + Codex) ADR-074 · Accepted
Vault Path & Payload Contract ADR-064 · Accepted
Kafka Consumer & Producer Durability Contract ADR-065 · Accepted
Feature-Flag Fail-Closed Policy ADR-066 · Accepted
Dual-Audience HTTP API Contract (LLM-First Conventions) ADR-067 · Proposed
Memory API in rag-gateway; mcp-bridge Becomes a Thin Proxy ADR-068 · Proposed
LLM-Facing Diagnostics API (Curated, Read-Only, sre-brain) ADR-069 · Proposed
Config-Keyed Client Singletons ADR-070 · Accepted
Graph Node Lifecycle & Event-Loop Offloading Contract ADR-071 · Accepted
Co-authored-by is a Lie: Cryptographic Provenance for AI Coding Agents ARTICLE · 11 minutes
Deprecate 99-Meta As Live Retrieval And Workflow Surface ADR-062 · Accepted
Developer-Agent Provenance & Cryptographic Attribution ADR-061 · Accepted
Rootweaver Deployment Contracts for Baked Images and TEI-Safe Embedding ADR-058 · Accepted
Memory-Writer Asynchronous Embedding (Decouple from Hot Path) ADR-059 · Accepted
Dynamic Test Impact Analysis for CI (selective per-package testing) ADR-056 · Proposed
MCP Bridge Efficiency: Composite Tools over Full Code-Execution Migration ADR-057 · Accepted
Ten bugs, two-tenths of a point — the weekend search got measurably better Season 3, Episode 7 · 16 minutes
Data-Store Backup Strategy & the Qdrant Snapshot-Disruption Tradeoff ADR-055 · Accepted
Dedicated CPU Embedding Worker Node (rduffy-mini) ADR-054 · Accepted
Personal Industry Radar Pipeline — Native Workspace Implementation ADR-050 · Accepted
Observability Data Routing — Multi-Backend Log/Trace/Metric Strategy ADR-053 · Accepted
Benchmark Agentic Metrics and Statistical Improvements ADR-052 · Accepted
Twenty-six hours, twelve tickets — and the audit that started everything else Season 3, Episode 6 · 25 minutes
Radar Feed Acquisition — Tiered Hierarchy for Sites Without Native RSS ADR-051 · Proposed
Egress Control Strategy — Forward Proxy Now, CNI Migration Later ADR-048 · Proposed
Personal Industry Radar Pipeline — Precis + Karpathy LLM Wiki + Claude Code Subscription Synthesis ADR-049 · Superseded-in-part-by-ADR-050
Memory-System Embedder Architecture — Dimension + Deployment Decision ADR-046 · accepted-with-caveats
Memory Eval Methodology v2 — Engineering Memory Focus ADR-045 · proposed
Memory System Derived-Data GDPR Position ADR-043 · Proposed
The Second Graph Comes Online: Activation, Communities, and an AI Reversal Season 3, Episode 5 · 22 minutes
Repository Structure for AI-Assisted Development ADR-030 · Accepted
The Refusal Gate: Teaching a Bot to Say I Don't Know Season 3, Episode 4 · 20 minutes
Daily Journal Consolidator External API Dependency ADR-037 · superseded
Daily Consolidator Reads JSONL Directly + Streaming Synth Model Swap ADR-038 · accepted
KEDA-Based GPU Swap Controller for Embed/Inference Coexistence ADR-029 · Accepted
Q2 2026 Retrieval, Inference & Evaluation Enhancement Program ADR-035 · Proposed
Vault Structured Index in Postgres ADR-036 · Accepted
Data Source Connector Architecture ADR-034 · Proposed
Agent Lifecycle Patterns — Lessons from Claude Code Source Leak ADR-026 · Proposed
MCP transport: moving from SSE to Streamable HTTP Design note · historical
NER System Evaluation — Span-Based Extraction ADR-040 · accepted
Measuring search quality with a repeatable benchmark Design note · historical
Trading Intelligence Platform Architecture ADR-025 · Proposed
Enterprise-Grade CodeGraph Platform ADR-020 · proposed
GitLab Flow Branching and Release Strategy ADR-024 · Proposed
HLD-Aware Architecture Review Pipeline ADR-023 · proposed
Local LLM Hybrid Routing — From Claude Code Augmentation to Full Replacement ADR-022 · accepted
Refusal gate: answer only when the evidence supports it Design note · historical
Short-Form Video Generation Pipeline ADR-033 · proposed
NL-to-Cypher LLM Backend Selection ADR-042 · accepted
Modular Architecture Documentation ADR-016 · Accepted
Graphiti + FalkorDB Knowledge Graph Integration Legacy ADR · Accepted
Matrix Accessibility Layer Architecture Legacy ADR · Proposed
MCP Bridge Authentication Strategy Legacy ADR · proposed
Streaming Journals: Kafka Meets LLMs Season 3, Episode 2
The Day Everything Got Sealed Season 3, Episode 3
Journal Pipeline Quality Overhaul — Model Selection & E2A Architecture ADR-012 · Proposed
Secret Management Strategy for Rootweaver Platform Legacy ADR · accepted
Journal Consolidator LLM Generation Tuning ADR-011 · accepted
K3s Pods Dashboard Architecture ADR-041 · accepted
LinkedIn Content Automation Pipeline ADR-001 · accepted
Journal LLM Context Handling for Large Sessions ADR-031 · Withdrawn-Duplicate-Of-ADR-032
Journal LLM Context Handling for Large Sessions ADR-032 · Proposed
Streaming Journal System Architecture ADR-010 · accepted
Anthropic GitHub Tools Integration ADR-039 · proposed
Anthropic GitHub Tools Integration Legacy ADR · proposed
Trunk-Based Development with Feature Flags Legacy ADR · Proposed
Unified Classification System Architecture Legacy ADR · Proposed
From Interview to Implementation: Perfect 100/100 Anthropic Alignment Season 2, Episode 5 · 42 min
Three Systems, One Weekend: The Parallel Testing Chronicle Season 2, Episode 4 · 18 min
54 Minutes to Production: Six Systems in One Day Season 2, Episode 3 · 28 min
48 Hours Later: From Validation to Velocity Season 2, Episode 2 · 21 minutes
The Validation Chronicles: 5 Hours of Truth Season 2, Episode 1 · 12 minutes
Day Zero: The ConvoCanvas Vision Season 1, Episode 1 · 7 minutes
Building the Foundation: Vault Creation to MVP Season 1, Episode 2 · 8 minutes
The AI Awakening: Breaking Free from Context Limits Season 1, Episode 3 · 8 minutes
ChromaDB Weekend: From 504 to 24,916 Documents Season 1, Episode 4 · 9 minutes
The Migration Question: When K3s Meets Reality Season 1, Episode 5 · 10 minutes
When Everything Crashes: The K3s Resurrection Season 1, Episode 6 · 9 minutes
Teaching the System to Document Itself: Automated Architecture Diagrams Season 1, Episode 7 · 8 minutes
Teaching the System to Blog About Itself Season 1, Episode 8 · 5 minutes
K3s on the home platform Design note · historical
No public platform notes yet. New writing and decisions are available in the other views.
Publication dates and public build notes. Journal entries describe earlier work. “Merged” does not confirm a live deployment.
The project behind the writing
Rootweaver
My independent AI engineering platform. Coding tools, persistent memory, searchable knowledge and code relationships, connected with signed provenance.
I design and operate it with AI-assisted development.