● Field notes — Amsterdam
The Sovereign Stack
Engineering autonomous AI from first principles. Architecture decisions, postmortems, benchmarks — the unglamorous infrastructure choices that decide whether AI products actually ship.
Detecting Embedding Drift: A Statistical Gate Before Sovereign RAG
Embedding drift silently degrades RAG quality. Learn to build a statistical gate that flags stale vectors before inference, using only your existing stack.
Postgres as Autonomous Memory: Crash Recovery for Vector Indexing
Vector indexing at scale is a crash-prone pipeline. Here's how to treat Postgres as the durable backbone of agent memory, with checkpoints and idempotent recovery.
NMI Watchdog Resets on 4-GPU Bare Metal: Tracing IRQ Storms to PCIe AER Errors
Hardware resets on 4-GPU nodes traced to NMI watchdog timeouts caused by IRQ storms from PCIe AER errors. A methodical debug path from dmesg to firmware settings.
Deterministic Replay of Agent Decisions: Beyond Temperature 0
Temperature 0 alone doesn't guarantee reproducible agent runs. Hidden state traps in sampling, kernels, and infrastructure can silently break determinism. Here's how to design for replay.
Ext4 vs XFS for Agent Memory Journals: A 90-Day Corruption Study
A 90-day crash-recovery benchmark on NVMe comparing ext4 and XFS for agent memory journals. Findings on corruption rates, recovery times, and practical implications for autonomous systems.
Pinpointing Agent State Corruption with Postgres xmin and LSN
When an agent's state goes bad, you need more than a hunch. Learn to use Postgres xmin and LSN to trace every write back to the exact commit that corrupted it.
Provenance Chains for Every Token: Linking Inference Logs to Training Data Lineage
How to build token-level provenance chains connecting inference logs to training data lineage, and why the EU AI Act does not actually require it.
WireGuard Mesh Pitfalls: MTU, Handshake Storms, and the 3am Timeout
WireGuard meshes look simple until they aren't. Here are the failure modes we've hit—MTU mismatches, handshake storms, and the dreaded 3am timeout—and how to diagnose them.
The 8-GPU NUMA Trap: CPU-GPU Affinity and Multi-Node Inference
Multi-GPU inference often stalls on NUMA mismatches. Learn how CPU-GPU affinity, PCIe topology, and memory placement shape latency—and how to diagnose it.
Why We Ditched Kafka for Postgres LISTEN/NOTIFY
We replaced Kafka with Postgres LISTEN/NOTIFY for agent event streaming. Simpler ops, fewer moving parts, and enough throughput for our scale.
Asynchronous Write-Through: Solving Cross-Store Conflicts Without Two-Phase Commit
When your AI stack spans Postgres and a vector store, two-phase commit is a trap. Async write-through with idempotent retries gives you consistency without the latency hit.
Crypto-Graph Replay: Verifying Agent State Transitions with Merkle Trees in Postgres
Learn how to verify agent state transitions using Merkle trees in Postgres, enabling deterministic replay and auditability without blockchain.
A GPU That Died at 92°C: Thermal Postmortem of a 4-Day Autonomous Agent Run
A 4-day autonomous agent run killed a GPU at 92°C. Here's the thermal postmortem and what to change for long-running inference.
fsync on Every Token: A Power-Loss Journal for GPU Inference
A write-ahead journal for token streams survives sudden power loss. Here's the architecture, the tradeoffs, and why fsync discipline beats buffering.
Sub-5ms Recall: Hot RAM, Warm SSD, Cold Tape for Agent Memory
Agentic systems need fast semantic recall without burning cash. A hot-warm-cold memory hierarchy—RAM, SSD, tape—keeps latency low and costs sane.
The 8-Hour Silent CUDA Leak: Tracing Attention-Mask Fragmentation in a Swarm
A swarm of autonomous agents slowly leaked VRAM over 8 hours due to attention-mask fragmentation. Here's how we traced it and fixed it.
Property-Based Chaos: Verifying CRDT Action Registers Without a Raft Leader
Raft gives you linearizability, but at the cost of a leader. For agent swarms, CRDTs offer an alternative—if you can prove they actually converge. Here's a property-based testing methodology that catches the bugs Jepsen
In-Process KV with WAL: Rethinking Agent State Sync
Moving agent state out of Redis into an in-process KV store with a write-ahead log simplified our stack and cut sync overhead. Here's the reasoning and a reference implementation.
Silent CUDA Leak: How Attention Masks Took Down Three Agents
A deep dive into a CUDA memory leak caused by attention masks in transformer inference, and how we traced it through agent crashes after eight hours.
fsync on Every Turn: The Real Cost of Durable Agent Memory
Durable agent memory means fsync. But fsync on every turn can tank throughput. Here's how to reason about the tradeoff and design for power-loss safety without crippling your agent.
Encrypting Model Weights at Rest Without Losing Inference Speed: A Hardware-Backed Approach
Model weights are your IP. Encrypting them at rest usually kills inference speed. Here's how to use hardware-backed keys and a decryption cache to get both.
Deterministic replay of LLM inference with temperature 0 and seed: what breaks in practice
Temperature 0 and a fixed seed don't guarantee identical outputs. Here's what actually breaks determinism in production LLM serving and how to work around it.
Why We Log Prompt Hashes and Outputs, Not Raw Text
The EU AI Act demands logging for high-risk AI systems. But logging raw prompts is a privacy and security nightmare. Here's how we do it with hashes and why it's better.
Three-Headed Memory: Resolving Cross-Store Conflicts for Agent Recall
Autonomous agents need more than one memory store. Here's how to reconcile vector, graph, and relational views without losing consistency.
Bare-Metal Inference Without Aircon: GPUs at 85°C and a Fan
Running GPUs without air conditioning is feasible if you accept high temperatures and manage thermal throttling with software. This article covers the practical engineering behind it.
Refusing to Answer: Calibrating Confidence Scores to Cut Hallucination
Calibrating confidence scores lets LLM systems refuse low-confidence answers, reducing hallucinations. A practical guide to thresholds, calibration, and guardrails.
GPU Firmware Bit Flips: How ECC Memory Saved Our Swarm from Silent Corruption
A single bit flip in GPU memory can corrupt weights and derail an entire agent swarm. Here's how ECC memory and vigilant monitoring saved our inference cluster from silent failure.
Raft for Agent Consensus: When Leaderless Coordination Beats Central Orchestration
Central orchestrators create single points of failure and coordination bottlenecks. Here's how we used Raft to build a self-organizing agent swarm that survives node loss and scales horizontally.
VRAM Fragmentation as a First-Class Failure: A Buddy Allocator for Inference Workloads
GPU memory fragmentation silently kills inference reliability. Here's how to design a buddy allocator that prevents it.
When Three Stores Disagree: A Conflict-Resolution Protocol for Agent Recall
Relational, vector, and graph stores drift apart in agent memory. Here's a deterministic protocol to detect and reconcile conflicts.
Deterministic Replay of Agent Swarms: A Clock-Free Event Log
Stop chasing flaky agent behavior. A clock-free event log gives you deterministic replay of multi-agent runs, making chaos reproducible.
Pseudonymizing Activations at the Tokenizer: EU AI Act Compliance Pattern
Learn how to pseudonymize model activations at the tokenizer level to reduce personal data exposure and align with EU AI Act requirements for transparency and privacy.
GPU Clock Throttling: Why Inference Latency Doubled at Hour 3
Sustained GPU inference can silently throttle clocks, doubling latency hours into a run. Here's how to detect, diagnose, and design around it.
Leaderless Swarm Coordination: CRDT Action Registers Over Raft
Raft gives you a leader; a swarm of autonomous agents often can't afford one. CRDT-based action registers let agents coordinate without consensus, at the cost of eventual consistency.
PCIe Lane Contention: Why Our 8-GPU Swarm Ran 30% Slower
We built an 8-GPU inference swarm, but it ran 30% slower than expected. The culprit wasn't software—it was PCIe lane contention. Here's how we diagnosed and fixed it.
Inline PII Stripping at 10k req/s: A Wire-Speed De-identification Proxy for Sovereign Inference
Building a transparent proxy that strips PII from LLM requests at 10k req/s with sub-millisecond overhead. A practical guide to sovereign inference.
Partial Rollback vs Full Replay: Cost Model for Poisoned Agent State
Agent state gets poisoned. Should you roll back a few steps or replay the whole run? Here's a cost model to decide.
Seeding Chaos: Property-Based Testing for Agent Consensus Under Network Partitions
How to use property-based testing with random network partitions to verify agent consensus protocols. A methodology for deterministic chaos in distributed AI systems.
Tiered Precision on a 4090: Why Per-Layer Quantization Beats Per-Model
Quantizing an entire model to a single precision wastes VRAM and hurts accuracy. Per-layer quantization assigns different bit widths to different layers, balancing memory and quality.
IPv6 Link-Local Only: The Sovereign Mesh Without DHCP or NAT
A deep dive into running a sovereign mesh on IPv6 link-local addresses alone—no DHCP, no NAT, just scoped addressing and deliberate routing.
Key Rotation Without Downtime: Re-Encrypting Vector Data In Place
Rotating encryption keys for a 2TB vector store without downtime is a hard problem. This post breaks down the tradeoffs and a concrete approach using envelope encryption, staged re-encryption, and versioned keys.
Three-Tier Agent Memory: When Hot RAM, Warm SSD, and Cold Tape Save a Run
A practical three-tier memory hierarchy for agent swarms: hot RAM for the working set, warm SSD for recent runs, cold tape for audit. When each tier earns its keep.
A 3-Second Power Blip Wiped GPU Buffers: The Journal That Fixed It
A brief power interruption can silently corrupt GPU buffers. A write-ahead journal for inference state turns a 3-second blip into a recoverable hiccup.
Priority Inversion in Agent Swarms: When a Low-Priority Task Stole the GPU
Low-priority agents can starve critical paths by hogging shared GPUs. Learn how priority inversion manifests in agent swarms and how to design around it.
Timeouts, Not Knowledge: When Legal-RAG Evaluators Measure Latency
A legal-RAG QA system scored poorly not because of retrieval or generation quality, but because the evaluator timed out waiting for deterministic citations. A post-mortem on nginx proxy limits and latency budgets.
Stale Telemetry Poisoned Our Agent Loop: The 24-Hour Freshness Gate
Lifetime telemetry skewed our agent's self-improvement proposals. We added a 24-hour freshness gate to keep the loop honest.
Prompt Injection via Vector Store: How a Poisoned Document Hijacked 40% of Our Agents
A single poisoned document in our vector store redirected 40% of our agents. Here's how it happened and how we hardened our RAG pipeline.
Event-sourced agent actions: why we log to postgres before the LLM even starts
Logging agent actions to Postgres before the LLM runs gives you durability, replayability, and auditability. Here's why we do it.
Embedding Drift: Detecting Silent Vector Corpus Decay
Your vector corpus decays silently as models and data shift. Here's a concrete methodology to detect drift before it breaks your RAG pipeline.
Write-Once Audit Logs on NVMe: Surviving Power Loss and Regulator Scrutiny
Build tamper-evident audit logs on NVMe SSDs using block-level write-once semantics, fs-verity, and power-loss-safe designs that satisfy EU AI Act requirements.