Pratyush Jena

I am an ML engineer with 8+ years of experience taking machine learning systems from problem framing all the way to production. I am based in India and have been working with distributed teams for the past few years.

Most of my recent work has been around LLM systems in production — retrieval pipelines, agent tooling, evaluation harnesses, and the unglamorous reliability work that turns a promising demo into something a business can depend on. I built a question-answering system over an internal knowledge base of around 2 lakh documents, bringing p95 latency down from 4.2 seconds to under 900 ms and taking eval-set accuracy from 68% to 89% by reworking chunking and adding a reranking layer. When plain vector search started breaking on multi-hop questions, I moved parts of it to a GraphRAG setup, building an entity and relationship graph so the system could follow connections across documents instead of just matching similar chunks. More recently I have been building MCP servers to give our agents clean, typed access to internal tools, which removed a lot of brittle glue code. My takeaway is that the model is rarely the hard part. It is retrieval quality, evaluation you can trust, and failure modes that only appear at scale.

This sits on a foundation of classical ML and data science. I have worked extensively on ranking and forecasting — demand forecasting that reduced inventory holding cost by roughly 12%, and a churn model still running in production three years later. I have also worked on NLP and computer vision, including an information extraction pipeline for scanned invoices. I prefer picking the boring model that works over the interesting one that does not.

I own the deployment side rather than handing models off. On AWS I have worked with SageMaker, EKS, Lambda, S3 and Bedrock, and I have also used Vertex AI and BigQuery on GCP. Around this I set up CI/CD with GitHub Actions, infrastructure as code with Terraform, and monitoring for system health and data drift. Moving batch inference to spot instances and right-sizing GPU nodes brought our monthly inference spend down by close to 35%. And I am the person on call when something breaks at 3 am.

Stack: Python, PyTorch, scikit-learn, LangChain, LangGraph, MCP, Neo4j, FastAPI, Postgres and pgvector, Airflow, Docker, Kubernetes, Terraform, AWS, GCP.

I am looking for a full-time remote role where I can own a problem area end to end, preferably with a team shipping AI products to real users rather than running pilots indefinitely. I am comfortable with async work and can overlap 4 to 5 hours with US Eastern or European time zones.