Architecture decisions. Constraint logs. What actually shipped.
Each post documents what the system required, what failed under load, and what the final design cost in tradeoffs. No opinion pieces. No trend commentary.
AI Systems · Embedded / ESP32 · Distributed Systems · Blockchain · Kafka Streaming · LLM Engineering · FreeRTOS · Kubernetes · RAG · Rust · Docker · Failure Analysis
Triage before you read
Documented decisions, reproducible results
Running inference on ESP32: memory constraints and what breaks first
Kafka at real-time scale: three pipeline designs before one held
Consumer group rebalancing, partition lag under burst load, and the offset commit strategy that finally matched the throughput requirement.
Quantized model weights, IRAM allocation limits, and the two architecture choices that forced a full redesign at the 80% mark.
RAG retrieval failures: when vector search returns the wrong context
Solidity reentrancy: the audit finding and the fix that introduced a new bug
Embedding model mismatch, chunking strategy errors, and the re-ranking layer that reduced hallucination rate from 31% to under 4%.
A checks-effects-interactions violation, the guard pattern applied incorrectly, and the second audit pass that caught the regression.
The Research Lab documents what the posts can't fit
Also in the lab: AI pipeline diagrams, engineering notebooks, and architecture decision records.
Experimental pipelines, architecture timelines, and failure logs that require more than a post format — organized for engineers who need the full constraint record.