Aniruddha Mandal

I build production-grade ML and LLM systems with a strong bias toward reliability, evaluation, and latency. As a founding engineer, I architected a multi-agent AI runtime service that powers end-to-end message classification and callout workflows.

I designed a sequential, context-grounded callout extraction pipeline using entity relationship graphs and entity resolution, reducing output fluctuation caused by batch processing. I also built real-time and batch enrichment flows that process Slack/Gmail messages through multi-stage LLM pipelines with tracing and dynamic prompt management.

I set up evaluation and quality gates with 27 graders across multiple agent types, automated Slack failure alerts, daily quality dashboards, and CI quality checks. To improve performance and cost, I implemented an end-to-end SLM fine-tuning pipeline (QLoRA) that brought classification latency from seconds to sub-second while cutting inference costs significantly.

Beyond classification systems, I’ve led AI research and engineering for conversational models—creating large synthetic datasets, training models benchmarked on EQBench3, and building personalization foundations. I’ve also delivered hybrid RAG architectures and knowledge-graph-driven solutions for clinical diagnostics and legal contract QA, partnering closely with engineering teams to ship end-to-end backends.