FastMCP Agent Mail: RBAC Tokens vs Anonymous Access, and the 403 Errors in Between
Why a FastMCP agent mail server that works anonymously in dev returns 403 behind TLS ingress, and how to wire bearer tokens without leaking them.
#ai-agents
23 posts
Hybrid Retrieval v2: Qwen Embeddings, BM25, and RRF with a FastEmbed Reranker
Dense vectors miss exact identifiers. Here's the hybrid Qdrant setup I run: Qwen3 embeddings, BM25 sparse vectors, RRF fusion, and a CPU FastEmbed reranker.
We Almost Deployed a Temporal Knowledge Graph. The Eval Said No.
A temporal knowledge graph passed every static RAG test and still failed. The gap was temporal reasoning, and one eval caught it before prod did.
I Benchmarked My Homelab Memory Stack: Hybrid Search + Local Reranker Took LoCoMo from 63% to 80%
How adding BM25 hybrid search and a local BGE reranker to a vector-only agent memory stack moved LoCoMo accuracy from 63% to 80%, and what it cost in latency.
The Write Policy Is the Hard Part: Promotion Pipelines for Agent Memory
Storing agent memory is easy. Deciding what earns a permanent write, and keeping the write-path alive through RBAC and network policy, is the real work.
Moving Scheduled LLM Curation from Cloud APIs to Local Models
A comparison of cloud APIs vs local models for scheduled agent memory curation, and the headless-cron gotchas nobody warns you about.
Silent Drift: Why Re-Embedding Only on Count Changes Rots Your Semantic Index
Count-based re-embedding leaves stale vectors pointing at edited text. Here's how semantic drift creeps in and how hash-based validation stops it.
Your Vector DB Snapshots Are Landing on the Same Disk That Will Fail
Qdrant snapshots default to the same volume as your data. One disk failure takes both, and even a surviving snapshot can fail restore with a 400 error.
Eviction Without Deletion: Running an ACT-R Decay Policy for Agent Memory
Stop deleting agent memories to fight vector-store noise. An ACT-R decay policy lowers activation instead, so cold context can re-activate later.
Langfuse for LLM Observability: Tracing Agent Calls Instead of Guessing
Using Langfuse to trace multi-step agent workflows, replace custom eval logic, and consolidate LLM observability into one tool that actually scales.
Agent Glass-Break Patterns: Controlled Escalation for Production
How to implement controlled escalation for AI agents using safeBins and network-level constraints to prevent production catastrophes.
When Agents Should Stop: Designing Safety Boundaries That Work
Stop conditions are a feature, not a failure state. Budget caps, progress checks, and exit contracts that make autonomous agents quit cleanly.
Building Agent Skills: A Pattern for Discoverable Capabilities
Moving beyond basic tool-calling to a composable skill pattern that makes agent capabilities discoverable and reusable across different frameworks.
Privacy-Routed LLM Inference: Keeping Sensitive Data Out of the Cloud
How to build a routing layer for AI agents that ensures sensitive data stays on local hardware while leveraging cloud LLMs for non-private tasks.
Cognitive Memory for Agents: Vector Search vs Activation-Based Recall
Comparing vector databases and activation-based memory for AI agents. Trade-offs in latency, scale, and interpretability.
Three-Layer Safety for Autonomous Agents: Stopping the Infinite Loop
Moving beyond prompt engineering to implement token-level schema enforcement, pre-execution gates, and shell-safe execution pipelines for AI agents.
Self-Improving AI Infrastructure: How Your Homelab Wiki Updates Itself
How to automate your homelab wiki with self-improving AI infrastructure
The 6-Layer Memory Architecture I Run for Claude Code
Open-sourcing the memory system behind my Claude Code setup: CLAUDE.md, path-scoped rules, wiki, vector search, cognitive memory. With the mistakes.
Building Karpathy's LLM Wiki: A Production Homelab Implementation
Implementing Karpathy's LLM Wiki in a homelab with real-world lessons and gotchas
Agent Credential Management: Two-Tier Service Accounts for Secure AI Agent Workflows
Managing agent credentials with two-tier service accounts: a secure approach for AI agent orchestration
NVIDIA Container Toolkit: Why the Default Runtime Matters
Fixing default runtime misconfigurations in NVIDIA Container Toolkit for GPU workloads
Building MCP Servers with FastMCP: Stop Writing Boilerplate, Start Writing Tools
FastMCP makes building Model Context Protocol servers feel like FastAPI. Here's how to go from zero to a working MCP server in under an hour.
Multi-Agent AI Systems: Architecture Patterns That Actually Work
A practical guide to designing multi-agent AI systems — orchestrator patterns, trust boundaries, and the tradeoffs I learned running agents in production.