Headline

Lead Data Scientist — Agentic AI & LLM Platforms | LangGraph · RAG · MLOps | I build the runtime layer under production agents | arXiv-published | M.Tech AI, IIT Jammu

About

I build the layer underneath agents. Most teams can get an agent working in a notebook. Getting it to run for many tenants, at sane cost, without a deploy every time the graph changes — that is the part I do. At Ada I replaced a monolithic agent runtime with a multi-tenant platform on LangGraph: Kubernetes runtime pools that scale per tenant, a graph cache that removes rebuild cost on every invocation, and a control plane that turned tenant config and rollout into something declarative instead of a release. Nine years in, that is the shape of my work — taking a model from something that demos well to something that survives production. What I do: → Agentic platforms — LangGraph, multi-tenant runtimes, graph orchestration, human-in-the-loop patterns → LLM applications — RAG pipelines and evaluation, open-weight models (Mixtral 8x7B, Gemma, Phi-2) fine-tuned and served on SageMaker, conversational AI → MLOps — CI/CD on AWS, DVC-driven retraining, post-deployment monitoring, continual learning → Leading teams — mentoring data scientists, prioritising roadmaps, turning model output into decisions stakeholders actually act on Two things I am glad I did: halved model deployment time by rebuilding CI/CD on AWS, and published "Enhancing Financial VQA in Vision Language Models using Intermediate Structured Representations" (arXiv:2501.04675) — fine-tuning DEPLOT on 50,000 charts so a chart becomes a structured table before the LLM ever reasons over it. Currently finishing an M.Tech in AI at IIT Jammu alongside the day job. Domains: telecom · fintech · healthcare data · real estate Stack: Python · LangGraph · LangChain · RAG · Mixtral / Gemma / Phi-2 · Transformers · TensorFlow · RASA · FastAPI · AWS SageMaker · Kubernetes · DVC · Azure · GCP · SQL Bangalore, open to remote across India. rajeshthakur1r@gmail.com

Location

Bengaluru

AI stack

aistudio

anthropic apis

claude code

claude

deepseek

hugging face

Work experience

Lead Data Scientist — Agentic AI & LLM Platforms at ADA, Sep 2025 - Present · 1 yr 1 mo, Bengaluru, I own the platform layer our agents run on. - Designed and shipped a multi-tenant agent platform on LangGraph, replacing a monolithic runtime driven by static config. Kubernetes runtime pools let tenants scale independently; a graph cache removes rebuild cost on every invocation; a control plane made tenant config and rollout declarative, so new agent graphs ship as configuration rather than releases. - Built and evaluated the RAG and evaluation pipelines around it, plus the MLOps that keeps the whole thing live. - Led a movie recommendation engine using collaborative filtering and clustering — 40% lift in user engagement. - Mentor a team of data scientists on ML practice and predictive modelling; own project prioritisation against business goals. - Built the dashboards and narratives that get model output in front of non-technical stakeholders. LangGraph · Kubernetes · Python · RAG · LLMs · MLOps

Lead Data Scientist — Generative AI & NLP at Synechron, Sep 2023 - Apr 2025 · 1 yr 8 mos, Bengaluru, - Fine-tuned DEPLOT on a 50,000-chart dataset so financial charts convert into structured tables before an LLM reasons over them — published as "Enhancing Financial VQA in Vision Language Models using Intermediate Structured Representations" (arXiv:2501.04675). - Shipped custom chatbots and NLP models on Mixtral 8x7B, Gemma and Phi-2, served and optimised through AWS SageMaker. - Signature verification and MICR detection using YOLOv9 and LangChain. - Customer-persona identification and credit-card recommendation with Nexus 360 and Table LLM, on SQL-driven data models. - Built post-deployment monitoring, evaluation metrics and continual learning for scalable NLP pipelines. - Delivered agentic AI for data processing and an AI-driven call management system.

Data Scientist — NLP & MLOps at Resident boost(zuma), Feb 2022 - Apr 2025 · 3 yrs 3 mos, Bengaluru, - Built and ran production NLP pipelines for text classification, entity extraction and summarisation, deployed on SageMaker. - Established continuous NLU training with DVC Studio and CI/CD on AWS with post-deployment evaluation — 50% faster model deployment. - Served models through FastAPI and Flask endpoints; built dialogue managers for natural language generation. - Shipped a RASA-based FAQ bot that cut the on-call load for NLP/ML support.

Resident Boost at Resident Boost, Feb 2022 - Sep 2023 · 1 yr 8 mos, Applied state-of-the-art machine learning techniques to build scalable, high performance NLP solutions, tackling challenges such as text classification and entity extraction. Established a continuous training pipeline for NLU models with DVC Studio, and maintained high-performance production NLP pipelines. Built and executed CI/CD pipelines using AWS tools, ensuring smooth deployment and post-deployment evaluations. Deployed scalable NLP models on SageMaker and developed dialogue managers for efficient Natural Language Generation. Created and maintained multiple endpoints for NLP actions using FastAPI and Flask, supporting seamless customer communication. Designed advanced machine learning models for large-scale data using ANN, DNN, and regression algorithms. Implemented and deployed an FAQ bot with the RASA framework, enhancing on call support for NLP/ML models.

Software Engineer at NoBroker.com, Apr 2021 - Mar 2022 · 1 yr, Bangalore Urban, Karnataka, India

Education

Master of Technology - MTech, Artificial Intelligence, Indian Institute of Technology Jammu

Bachelor of Engineering - BE, Electrical, Electronics and Communications Engineering, Ragiv gandhi prodhyogki viswavidyalaya

Public links

github: https://github.com/RajeshThakur1

youtube.com: https://youtube.com/@pybron