Field notes on building GenAI systems that hold up in production, and the decisions behind them.
A practitioner's guide to choosing between prompting, RAG and fine-tuning, and to the parameter-efficient methods (PEFT, LoRA, QLoRA) that make fine-tuning practical.
Lessons from a RAG assistant that resolves about 65% of HR, IT and operations questions without a human, and why the hand-off matters as much as the answers.
Generating SQL is the easy part. Validation, orchestration and returning a chart instead of a table are what got 300+ users to stop filing report requests.
Accuracy, latency and cost all compound in a speech pipeline. Here is how we got a STT → LLM → TTS banking agent to 92% intent accuracy at 1.8 s p95 and ₹0.48 a turn.
Why a money-mule platform needs more than a good model: real-time scoring, a network view, cost-aware thresholds, and explanations investigators can act on.
Starting this blog to document the architecture decisions, failure modes, and lessons behind the enterprise GenAI systems I build.
