From research to production — I build ML systems that work in the real world.
Data analyst at Novartis Healthcare and ML researcher based in Hyderabad, India. I work across the full stack — from CUDA kernels and transformer architectures to CRM platforms and business analytics. BTech CSE (AI & ML), Vellore Institute of Technology Chennai, CGPA 8.78. Open to roles.
Distilled CLIP ViT-B/16 into EdgeNeXt-XX-Small (approximately 66x compression) for on-device crop disease detection across 38 disease classes. 3-stage curriculum with a text-prototype-anchored distillation loss. Targets real-world inference on resource-constrained agricultural hardware.
Stack: CLIP, PyTorch, Edge AI, Knowledge Distillation.
Co-authored position paper arguing translation — not symbolic inference — is the bottleneck in LLM+solver systems. Evidence from an 85-paper review, multi-scale probing of GPT-2 and DistilBERT, and Phi-2 and Qwen2.5-3B benchmarks on FOLIO with the Z3 SMT solver.
Stack: Position Paper, GPT-2, Phi-2, FOLIO, Z3 SMT.
Decoder-only transformer that allocates compute per-token via representational convergence gating — simple tokens early-exit, complex ones run full depth. 29% FLOPs reduction with matched perplexity on WikiText-103 across 13 architectural patches.
Stack: PyTorch, Transformers, NLP.
Led design and integration of Meta's NLLB model into a Bixby capsule for auto-detection and translation across 200+ languages at 98% accuracy. Hosted via Docker and Hugging Face during Samsung R&D internship. First author on findings published on arXiv (arXiv:2403.05982).
Stack: NLLB, Docker, Hugging Face, Bixby.
Natural-language analytics agent over business databases. Ask a question in plain English, get SQL, a chart, and a narrative summary in one shot. RAG retrieval over schema docs keeps hallucinations near zero.
Stack: LangChain, RAG, FastAPI, GPT-4o, Plotly.
Multi-agent LLM pipeline for pharma document intelligence — adverse-event extraction, compliance checks, and structured CRUD summaries from unstructured visit notes and Vault documents. Built around real pain points from the Novartis CRM workflow.
Stack: LLM Agents, Python, Pharma, NLP.
Production forecasting service using PatchTST for demand and KPI prediction. Exposes a REST endpoint: upload a CSV, get back forecasts with confidence bands and anomaly flags. Packaged with MLflow tracking and Docker-compose deployment.
Stack: PatchTST, MLflow, FastAPI, Docker.
PyTorch, TensorFlow, Keras, Scikit-learn, Hugging Face Transformers, Deep Learning, NLP, Computer Vision.
Python, C++, Docker, PySpark, CUDA.
SQL, MySQL, PostgreSQL, Snowflake, FAISS, DataIKU DSS, Alteryx, Power BI.
JavaScript, Data Structures and Algorithms, Object-Oriented Programming, Statistical Modeling.
Stakeholder Management, Business Requirement Structuring, Solution Design, KPI-driven Decision Making, Salesforce.
Building something at the intersection of ML and production? Let's talk.