Skills

What I actually use at work — with context on how and why.

Modeling & ML

My modeling work spans supervised ranking, classification, and generative tasks. At Myntra I work primarily with gradient-boosted trees and neural ranking models for CTR prediction on sparse ad auction data. At MIDAS Lab I fine-tune and evaluate vision-language models for annotation tasks and small LLM generation improvement.

PyTorchXGBoostLightGBMScikit-learnHugging Face TransformersLearning to RankONNX

Data Engineering

I build feature pipelines that are idempotent, testable, and reproducible. On the AdsRank team I re-architected snapshot-based pipelines on Databricks and PySpark, replacing ad-hoc scripts with versioned feature tables. I care about lineage and correctness first, throughput second.

DatabricksPySparkApache SparkDelta LakeSQLdbtPandas

Production Infrastructure

I've deployed inference APIs using FastAPI and Celery for async batch workloads, containerized with Docker, and monitored with Grafana/Prometheus. I'm comfortable with the DevOps side of ML — experiment tracking, model registries, and CI pipelines for data and model validation.

FastAPICeleryDockerKubernetesRedisMLflowGrafanaPrometheusGitHub Actions

LLMs & Multimodal AI

I work on LLM evaluation, prompting strategies, and lightweight fine-tuning. My research at MIDAS Lab focuses on improving small LLM generation quality with Mixture of Refinement Agents. I've also built multimodal pipelines using vision-language models (QWEN-2.5-VL, CLIP) for classification and anomaly detection tasks.

OpenAI APIAnthropic ClaudeQWEN-2.5-VLCLIPMCP (Model Context Protocol)LangChainStreamlit