I build things at the intersection of classical ML and Generative AI; not just wrapping an LLM API call, but fine-tuning models, designing RAG pipelines, and shipping the plumbing (APIs, dashboards, extensions) that makes a model actually usable.
- 💼 Currently a Forward Deployed Engineer Intern @ UnifyApps — a role blending presales and hands-on technical implementation across integrations, automations, data, and agentic AI/MCP workflows
- 🔭 Pursuing my M.Sc. in Data Science at NMIMS, Mumbai — CGPA 9.45 (9.27 → Sem 1, 9.82 → Sem 2)
- ✨ Recent focus: Generative AI in practice — fine-tuning transformers, building RAG pipelines over ChromaDB, and using LLMs as reasoning components inside larger systems rather than standalone chatbots
- 💬 Core stack: Python, SQL, Machine Learning, Deep Learning, Gen-AI, Data Visualization — with secondary experience in React.js, Next.js, and full-stack web development
- 🏗️ I like projects that go end-to-end: data ingestion → modeling → explainability → a real interface someone can actually use
- ⚡ Fun fact: there are more possible chess game variations than atoms in the Milky Way galaxy
Languages & Core
Machine Learning & NLP
Agentic AI / LLM Systems
Data Engineering & Cloud
Data Tools & MLOps
Deployment & Web
AI Forward Deployed Engineer Intern @ UnifyApps
|
End-to-end NLP & MLOps — YouTube comment sentiment analysis Fine-tuned Twitter-RoBERTa on 1M+ YouTube comments to classify sentiment on internet slang, emojis, and creator-audience language that traditional models miss. Took the full journey from Bag-of-Words → LightGBM+Optuna → transformer fine-tuning, benchmarking every stage. Result: 88% accuracy / 87.7 macro-F1 on a hand-curated real-world validation set, deployed via FastAPI on Hugging Face Spaces with a Chrome extension front end.
▶ Demo · 💻 Repo · 🤗 Model · 📖 API Docs |
Hybrid AI fraud & phishing detection system 🏆 CipherCop 2025 National Finalist Combines three detection pillars in parallel: a LightGBM model on 30+ URL/domain features (with SHAP explainability), a local Mistral 7B LLM for contextual content analysis, and a computer-vision brand-similarity engine (pHash/dHash + OCR) — fused into a single weighted verdict with adaptive thresholding. Result: 94.7% accuracy and 0.96 AUC-ROC across 28,500 URLs — a 32% improvement over single-model baselines. Full analysis in under 30 seconds, with a dedicated frontend, explainable AI-powered risk assessment, and decision support.
|
|
Multi-agent business intelligence assistant A LangGraph-orchestrated system where specialized agents (Planner, Router, Coding, Research, RAG, Critic, Compiler) collaborate with human-in-the-loop approval and self-corrective reflection to turn a business question into a validated, sourced report. Result: Parallel agent execution with a reflection/retry loop that catches low-confidence or contradictory outputs before they reach the final report.
|
Production-grade lakehouse on Databricks A medallion-architecture (Bronze/Silver/Gold) pipeline simulating a 5-location restaurant chain — dual-source ingestion (Azure SQL CDC + Event Hub streaming), star-schema modeling, and in-pipeline LLM sentiment/issue classification on customer reviews via Result: Two live AIBI dashboards (chain performance + review insights), incremental Gold materialized views, ~60% reduction in data processing overhead through optimized compute and incremental processing.
|
📌 More on my portfolio site
🥈 National Finalist — CipherCop 2025 Bureau of Police Research & Development (BPR&D) × Telangana Cyber Security Bureau (TGCSB) — one of 28 finalist teams selected from 360 entries in a national-level cybersecurity hackathon. Presented the "Expose.AI" phishing detection prototype (built on the Spot the Fake system) live in Hyderabad to a panel of senior IPS officials, cybersecurity experts, and academic leaders from the Indian School of Business (ISB).