Headline

Staff ML Engineer @ Clari

About

I’m a Data Scientist with experience in building end-to-end machine learning systems—from research and experimentation to scalable, production-ready solutions. My work spans across domains like NLP, GenAI, recommender systems, audio classification, and knowledge extraction, with a strong focus on real-world impact and product alignment. I’ve led projects involving fine-tuning and deploying large language models (LLMs), architecting RAG pipelines, implementing prompt engineering strategies (e.g., CoT, Text2SQL), and developing multi-agent GenAI systems. I’ve deployed solutions using tools like Hugging Face, PyTorch, vLLM, TGI, and built scalable inference systems with Kubernetes, Docker, and cloud platforms like AWS. Beyond model building, I take ownership of the full ML lifecycle—from data pipeline design (Spark, Kafka, Airflow) to deployment and monitoring—while mentoring team members and working closely with cross-functional stakeholders. I’m passionate about creating AI systems that are reliable, interpretable, and user-centric, and I enjoy solving ambiguous problems with practical, high-quality solutions. Let’s connect if you’re working on exciting problems in ML, GenAI, or AI product development!

Location

Bengaluru

AI stack

claude

hugging face

langgraph

langsmith

DSPy

vLLM

Work experience

Staff Data scientist at Clari, Sep 2025 - Present · 1 y, Bengaluru, Karnataka, India

Data Scientist III at Draup, Jul 2023 - Aug 2025 · 2 yrs 2 mos, Bengaluru, Karnataka, India, Led the design and development of core GenAI and ML infrastructure at Draup, taking ownership of foundational systems that power multiple downstream products — while also guiding and mentoring data scientists across the team on model design, experimentation, and deployment practices. - Built Curie, an in-house GenAI assistant orchestration framework handling complex platform queries through a combination of RAG, Text-to-SQL, and tool calling. The system leveraged LLMs, Chain-of-Thought prompting, and multi-agent logic to deliver accurate, data-driven responses across diverse data sources — a foundational architectural decision that standardized how the platform exposes AI-driven query capabilities. - Designed and fine-tuned a Mistral-7B-Instruct model for News Entity Relation Extraction using LoRA and AWQ 4-bit quantization, with Pydantic-style structured prompting for extracting entities and relationships from business news. Deployed via vLLM with paged attention for optimized inference, reducing extraction turnaround from months of manual effort to just a few hours. Synthetic training data was generated using ChatGPT to bootstrap the fine-tuning pipeline. - Designed and owned the JD Segregation system end to end — a multimodal MarkupLM (RoBERTa-based) pipeline fine-tuned on HTML xPaths and text to extract structured components from job descriptions. Deployed inference via Airflow DAGs processing ~1M incremental JDs daily, and ran bulk inference on ~500M historical JDs on a distributed Spark and GPU setup optimized with Kernel Fusion via Better Transformer. This system now serves as a foundational model for multiple downstream ML products. - Mentored and guided data scientists on model design, experimentation, and production deployment practices

Data Scientist II at Draup, Jul 2021 - Jul 2023 · 2 yrs 1 mo, Bengaluru, Karnataka, India, Worked across a suite of AI-powered features centered around intelligent user profiling, content understanding, and recommendation systems. - Fine-tuned a BART-large decoder-only model to generate abstractive summaries from course content, significantly reducing manual effort in publishing workflows. - Built a unified LLM Manager as a configurable wrapper over multiple LLM providers to standardize sync and async calls across teams using LiteLLM. - On the recommendation side, built two distinct systems — a skill recommendation engine using BGE-base embeddings with relevance scoring for core and soft skills, and a company recommendation system using multi-feature similarity across verticals and functional parameters, with rationale generation aligned to user-selected companies.

Machine Learning Engineer at HCL Technologies, Oct 2020 - Jul 2021 · 10 mos, Bengaluru, Karnataka, India, Built an NLP-powered document intelligence pipeline for Medmind as part of my work for Clarivate Analytics — a platform helping pharma companies track regulatory changes across drug-related policy documents. The system automated extraction of key insights from complex PDF documents using NLP techniques, with Tabula handling structured PDF parsing — all orchestrated through a Flask-Celery-RabbitMQ pipeline built for reliable async document processing at scale.

Data Scientist at Genisys Group, Mar 2019 - Oct 2020 · 1 yr 8 mos, Bengaluru Area, India, Led applied research and deployment of production-grade NLP and Audio ML systems for a government project. - Built an unsupervised audio preprocessing and CNN-based classification pipeline for speech, music, and noise detection — deployed on Kubernetes with TensorFlow Serving and Kafka for real-time inference - Built a real-time NER Engine using Spark-NLP with DL and CRF models on IOB-tagged data, deployed via Kafka-based MapR Streams for high-throughput streaming inference in production

Senior Software Engineer at Sopra Steria, Aug 2018 - Feb 2019 · 7 mos, Worked on applied NLP research across NMT and seq2seq architectures for natural language understanding, and built intelligent semantic search solutions beyond keyword matching. Mentored junior data scientists on NLP concepts, model experimentation, and production best practices.

Software Engineer at Sopra Steria, Jan 2017 - Aug 2018 · 1 yr 8 mos, Built Botify.kit — an internal platform that automated chatbot creation by running an NLP pipeline on raw QnA documents to extract intents and entities and deploy IBM Watson-powered bots.

Software Engineering Trainee at Sopra Steria, Jul 2016 - Jan 2017 · 7 mos, Built an HR Policy Chatbot using Node.js, Express.js, and IBM Watson — handling dataset curation, intent classification, and entity recognition to create a working conversation flow.

Education

Engineer’s Degree, Information Technology, Inderprastha Engineering College

AISSCE (CBSE), Physics, Chemistry, Maths, Careers India Jobs

Public links

github: https://github.com/umangkeshri

leetcode: https://leetcode.com/umangkeshri