Writing and technical notes

Engineering ideas, documented with context and constraints.

These outlines will become full articles on this site and DEV Community. They are grounded in implemented systems and research rather than generic tutorials.

CUDA · PyTorch

GPU-accelerating MSCRED with im2col and GEMM

Problem framing, model bottleneck, custom extension design, shared-memory tiling, benchmarks and limitations.

Outline ready
QLoRA · Reasoning

Fine-tuning Nemotron-3-Nano-30B for structured reasoning

Task taxonomy, compact synthetic traces, 4-bit adaptation, compatibility fixes and submission packaging.

Outline ready
RAG · Evaluation

Engineering an evaluation-driven enterprise RAG pipeline

Ingestion, query expansion, hybrid retrieval, reranking, reflective retries and retrieval metrics.

Outline ready
LLM quality

Why LLM evaluation needs more than a single score

Golden datasets, repeatability, LLM-as-judge, category-level regressions and quality gates.

Outline ready

Publishing plan

Original here. Distributed elsewhere.

The portfolio will remain the canonical home. Articles can then be cross-posted to DEV using the original URL, with shorter summaries shared on LinkedIn.