Senior ML Engineer
I build data-intensive applications and the pipelines underneath them — ingestion, training data, and the inference infrastructure that has to stay fast and affordable once a model reaches production. Ten years of it, mostly in Python on AWS, most recently on computer-vision and generative-AI systems for insurance.
I started in product engineering, moved into data engineering, and now work at the seam between the two — taking models that work in a notebook and making them survive production traffic. The problems I find most interesting are the unglamorous ones: why inference costs what it does, why a pipeline silently produced the wrong number, why a retrieval system returns the superseded version of the right document.
Currently at EXL Services, working on GPU-accelerated aerial-imagery inference and enterprise RAG pipelines for insurance risk and claims.
EXL Services · India
Own the deployment and development pipelines for computer-vision and generative-AI systems serving insurance risk and claims processing — from GPU inference optimisation through to production serving on AWS.
Tiger Analytics · India
Built the PySpark analytical ecosystem that unified first-party and third-party data into model-ready datasets for downstream analytics.
Mindtree · India
Built internal tooling, data-processing CLIs and Flask microservices, and shipped the first production ML model of my career.
High-throughput PyTorch inference for aerial imagery, used to assess property risk and process insurance claims.
End-to-end retrieval-augmented generation pipeline — ingestion, embedding generation, vector search and inference — for enterprise applications.
A data pipeline that derives structured flags from free-text claim notes, acting as the central feature store for predictive models.
A walkthrough of exporting PyTorch models to ONNX for deployment — what actually gets smaller, what doesn't, and the export gotchas worth knowing before you ship.
Retrieval quality, not generation quality, is where enterprise RAG systems fail. Notes from building ingestion, embedding and vector search for enterprise document corpora.
Third-party data changes shape without telling you. Notes on building ingestion that fails loudly at the boundary instead of silently three tables downstream.