Engineering notes
Short, practical writing on production systems, edge AI, robotics, networking, and agent reliability.
Vector Index Migrations Need Query-Level Recall Diffs
A rebuilt vector index can be complete, healthy, and fast while returning materially worse evidence. Safe migrations compare old and new retrieval results per query before traffic is promoted.
Watchdogs Need Progress Proofs, Not Process Heartbeats
A thread can keep sending heartbeats while useful work is stalled behind a deadlock, saturated queue, or broken dependency. Reliable watchdogs monitor bounded forward progress and preserve evidence before recovery.
Agent Executions Need Reproducible Environment Manifests
An agent transcript does not capture the repository, tools, permissions, services, and network state that shaped an execution. Reproducible environment manifests turn surprising runs into reviewable engineering evidence.
Robotics Command Paths Need Expiring Authority Leases
When autonomy, teleoperation, and maintenance tools can all command a robot, source priority is not enough. Expiring authority leases make ownership, handoff, and stale-command rejection explicit.
DMA Buffers Need Explicit Cache-Coherency Contracts
A DMA transfer can complete successfully while the CPU still reads stale bytes or a device receives old data. Reliable zero-copy pipelines make ownership, synchronization, and cache-coherency transitions explicit.
Edge AI Rollbacks Need Artifact Compatibility Contracts
Rolling back an edge-AI executable is not enough when models, TensorRT engines, calibration data, schemas, and caches evolved with it. Safe recovery treats the deployable unit as a compatibility-checked artifact set.
Packet Drops Need Provenance, Not One Counter
A total packet-drop counter proves that traffic disappeared but not where, why, or under which pressure. High-performance data planes need stage-specific reason codes and enough context to distinguish policy from overload and defects.
Edge Fleets Need Configuration Provenance
An edge device should be able to explain the exact effective configuration that produced its behavior. Provenance links every value to its source, precedence, version, and rollout so configuration drift becomes diagnosable evidence.
ROS2 QoS Is a Failure Contract, Not a Tuning Menu
ROS2 QoS settings decide what a robotics system does when messages are late, lost, replayed, or disconnected. Profiles should be derived from topic failure semantics and verified under fault injection, not copied from defaults.
Zero-Copy Pipelines Need Explicit Buffer Ownership
Removing memcpy calls does not make a pipeline safely zero-copy. High-throughput systems need an explicit ownership protocol for DMA buffers, including transfer points, lifetime rules, backpressure, and reuse evidence.
Bug Bounty Reports Should Ship as Reproduction Harnesses
A strong bug-bounty report does more than describe surprising behavior. It packages preconditions, bounded test cases, postcondition checks, and cleanup into a deterministic harness engineers can rerun while fixing the system.
Retrieval Quality Needs Release Gates, Not Demo Queries
A retrieval system should not be tuned by trying a few convincing prompts. A deterministic corpus, explicit relevance judgments, quality metrics, and latency thresholds turn retrieval changes into reviewable release decisions.