Senior Machine Learning Engineer (ML Platform & Frameworks)

• Lead fault-tolerant distributed inference and training infrastructure for the org's ML platform — GPU pipelines on Kubernetes engineered to survive node loss and partial failure, from hyperscaled daily ingestion through production model inference.
• Designed and led the org's online and offline inference systems: online inference on vLLM + Ray for rapid experimentation with LLMs such as Falcon-40B and Llama 2 70B; offline inference on Ray Data + PyTorch Lightning, introducing torch.compile(mode="max-autotune") in production as a compiler bridge — compiler-generated Triton kernels, operator fusion, and CUDA-graph capture.
• Modeled data pipelines as horizontally-scaling Source → Transform → Sink plugins with no pod-to-pod communication, KEDA-autoscaled on SQS queue depth.
• Leading development of a PyTorch foundation-model training framework adopted across the org for training generative models on billions of images and videos — a plugin-based strategy pattern over FSDP/FSDP2, Flash Attention 2/3, and context/distributed attention for DiT text-to-image and text-to-video training.
• Built the CLI-driven build and deployment system (BuildKit images on KEDA-scaled Kubernetes jobs, pushed to ECR), fronted by a Rust control plane and an MCP interface so AI agents can drive builds, deployments, and inference.
• Solely authored the org's fault-tolerant training-data governance and lineage subsystem in Rust: idempotent claim/receipt registration, immutable S3 evidence layouts, and manifest-after-data publication.
• Wrote the org's first data quality framework on Cerberus, used across enrichment pipelines to validate the correctness of feature-generation outputs.

Software Dev Engineer II (ML Platform)

Bengaluru, India

• Was part of team that worked on Feature Store and pipeline which feeds on-demand features to deployed ML models at production scale(4Bn rows, 10K QPS). Pipeline-supported multichannel ingestion, i.e. Spark, Flink and user files etc.
• Founding member of Forecasting and Correlation Platform which was considered by many teams to forecast concerned time series. These forecasts power critical scaling decisions across organizations in real time.
• Led DAQ, a tool used to scrape APIs at scale. Used to collect data for analysis/ model training at a scale of 15 M rows daily.

Software Development Engineer (Search Relevance)

Bengaluru, India

• Was responsible for improvements/inception of search intent models(CRF/Neural Network based), identifying error classes, coming up with solutions, and fixing them. These models power user search and discovery for millions every day.
• Implemented a FastText based query store classifier, which predicts the category of a tail query.
• Implemented the first workflow to automate training and auto-deployment of various search models in Flipkart. First was written using Luigi and later migrated to Airflow.
• Wrote a generic framework using Airflow which at runtime creates generic dags for different ML models and orchestrates their training to deployment flow, including data and model validations.
• Implemented large scale (4Bn+ datapoints) pipelines using Cascading/HDFS to extract data from user events and then transform it to be used for training these models.

Software Development Engineer

Bengaluru, India

• Worked on a component called Cyclops, an interface between Customer representatives and internal services.
• This service is live in all countries in which Groupon operates.

Software Engineer

Bengaluru, India

• Led engineering efforts on modules like Polarity based Arbitration, Multi-Cast Filtering, Structural Latency Breakdown, etc.