Writing about backend engineering and GenAI — what works in production, what doesn't, and the gaps between tutorials and real systems.
What I learned building a QLoRA → GPTQ → vLLM pipeline for Gemma 4 31B: a documented quantization bug that costs 16 accuracy points, dependency hell, and why inference speed is about memory bandwidth, not compute.
Prompt caching strategies, PII handling, security layers for AI systems, and how queue-based multiprocessing achieved a 40x speedup — practical patterns from PyCon India 2025.
Model Context Protocol architecture, multi-agent evolution patterns, and practical lessons from PyCon India 2025 on building AI systems that actually work in production.