Headline

Agentic AI Engineer | RLHF · LangGraph · RAG · LLMs | Building AI agents in production | Qdrant Star ⭐ | IEEE Published

About

Agentic AI Engineer building systems that actually run in production. At Sutherland, I architect RLHF pipelines for Text-to-SQL from first principles — preference data collection, custom reward models, PPO policy optimization — serving 10K+ daily transactions on AWS. I also built a multi-stage agentic inference system with quantized LLM serving (4-bit/8-bit) that cut latency 60% at F1 > 0.95. I'm a Qdrant Star: open-source contributor to vector search and RAG infrastructure. Built SQLRouter (agentic Text-to-SQL with semantic caching) and a sandboxed agent execution runtime from OS primitives. Research background: CVIT, IIIT Hyderabad (collaborated with Meta Reality Labs). Published at IEEE ICETITE and Springer ICCCN. IndoML 2022: 90.8% at IIT Gandhinagar. Stack: LangGraph · LangChain · RLHF · PPO · RAG · Pinecone · Qdrant · FastAPI · PyTorch · AWS · Docker Open to Agentic AI / LLM Engineering roles where AI is the product, not a bolt-on. DM me if you're building production agents and hitting retrieval or latency walls.

Location

Bengaluru

Work experience

Data Scientist at Sutherland, Aug 2025 - Present · 1 yr, Chennai, Tamil Nadu, India, Building production-grade AI systems at the intersection of LLMs, RLHF, and agentic inference. • Architected an end-to-end RLHF pipeline for Text-to-SQL from first principles — designed preference data collection, trained a custom reward model on human feedback, and applied PPO policy optimisation to steer SQL generation, serving 10K+ daily transactions on AWS with measurable intent alignment gains. • Engineered a multi-stage agentic inference system with quantized LLM serving (4-bit/8-bit) cutting latency by 60% at F1 > 0.95; designed a RAG layer over Pinecone for sub-200ms retrieval, wiring tool-calling, schema grounding, and safety validation into a single orchestration loop. Stack: PyTorch · HuggingFace · PPO · RLHF · Pinecone · AWS · SQL

Open Source Developer Advocate at Qdrant, Sep 2025 - Present · 11 mos, India, Qdrant Star contributor focused on vector search, RAG architecture, and real-world retrieval systems. • Built SQLRouter — a modular agentic Text-to-SQL system with semantic query caching over Qdrant, short-circuiting redundant LLM calls at scale across intent classification, schema retrieval, SQL generation, and safety validation. • Built Contract Voice Agent — a full-stack agentic RAG system for document risk analysis with clause-level embeddings, hybrid retrieval with pre-tagged risk signals, and a real-time conversational interface over FastAPI + React + WebSocket. Stack: Qdrant · RAG · LLMs · FastAPI · Vector Search · Embeddings Featured on : https://qdrant.tech/stars

Machine Learning Engineer at Quoppo, Oct 2024 - Apr 2025 · 7 mos, Pune District, Maharashtra, India, • Designed hybrid LSTM-MeshNet architecture for 3D positioning from scratch; diagnosed and resolved vanishing gradient failure modes via custom multi-term loss scheduling and adaptive LR warmup, achieving 30% accuracy improvement over SOTA baseline. • Rebuilt GAN training loop with redesigned discriminator and curriculum-based fine-tuning; reduced post-production editing time by 40% and established reproducible MLOps foundation (W&B + DVC + CI/CD) across a 3-engineer team. Stack: PyTorch · GANs · LSTM · Weights & Biases · DVC · GitHub Actions

Undergraduate Research Fellow at Centre for Visual Information Technology (CVIT), Mar 2024 - Aug 2024 · 6 mos, Hyderabad, Telangana, India, • Implemented advanced neural network algorithms to enhance OCR accuracy by 20% on low-resolution images. • Collaborated with Meta Reality Labs for experimental validations, achieving a breakthrough in image processing technology. • Conducted a comparative analysis between diffusion models and GANs to improve prediction accuracy by 15%.

NLP Intern at Textify AI (Backed by AngelList), Oct 2023 - Jan 2024 · 4 mos, Indore, Madhya Pradesh, India, • Developed and implemented a cutting-edge ML-based Ticket Qualification Platform using FastAPI, Langchain, Node.js, and MongoDB. • Achieved an impressive F1-Score of 99.8% for predicting ticket categories and 93.7% for severity levels with the NLP model developed. • Enhanced user experience by integrating FastAPI framework and Langchain algorithm to reduce translation errors and improve server response time.

Education

Bachelor of Engineering, Information Technology, Pune Institute of Computer Technology