A running record of AI systems built, experiments run, and architectures designed — from LLM pipelines to production inference engines.
How I reduced p99 inference latency by 63% on a 70B parameter model through continuous batching, paged attention tuning, and a custom speculative decoding pipeline. Full architecture walkthrough, benchmarks, and the lessons that didn't make it into the paper.
Read PostCombining dense vector retrieval with BM25 sparse search and a cross-encoder re-ranker to push retrieval recall above 94% on domain-specific corpora.
Read PostWhat actually breaks in multi-agent systems at scale — loop detection, tool call reliability, memory contention, and the surprising cost of over-planning.
Read PostDesigning an autoscaling strategy on AWS that handles 10× traffic spikes without pre-warming costs eating your margin. Spot instances, queue depth triggers, and warm pool tuning.
Read PostFine-tuning a 13B model to 99% of full fine-tune quality on a single 24GB GPU — quantisation choices, rank selection, and gradient accumulation tricks.
Read PostAn empirical analysis of CoT reasoning quality vs. token budget constraints — and a structured prompt compression technique that preserves 88% of reasoning depth.
Read PostBenchmarking knowledge-graph-augmented retrieval against standard vector search on 10,000-document legal corpora — where graph wins, where it doesn't.
Read PostBuilding deterministic, type-safe tool interfaces for LLM agents — schema design, retry logic, and observability patterns that make debugging tractable.
Read PostEnd-to-end streaming architecture from GPU inference to browser — backpressure handling, connection pooling, and graceful degradation under load.
Read Post