BTech in Artificial Intelligence & Data Science · 2 years as an AI Engineer. I build production-ready systems: multi-agent LLM pipelines, generative AI (LoRA training, diffusion, video generation), and full-stack AI products. Based in India.
🤖 Agentic & LLM Systems — Multi-agent pipelines (LangGraph), RAG with structured evidence passing, fine-tuning with SFT + GRPO on Qwen and Llama models
🎨 Generative AI & Vision — LoRA training pipelines (Kohya_ss → SD / FLUX.1), text-to-video (Wan2.2), computer vision (CLIP · RetinaFace · YOLOv5)
🏗️ Full-Stack AI Products — FastAPI + Next.js + PostgreSQL, deployed apps on Render, deterministic safety layers over LLM outputs
📊 Data & Financial ML — Time series forecasting, crypto/retail analytics on Snowflake, clustering and EDA pipelines
LangGraph OpenAI API Anthropic API Gemini API DeepSeek Llama 3.x Qwen Hugging Face vector database
Stable Diffusion FLUX.1-dev Wan2.2 LoRA · Kohya_ss CatVTON CLIP AMD ROCm
| Project | What it does | Stack |
|---|---|---|
| Multi-Agent RAG | Three-agent literary consistency pipeline: Llama 3.3-70B decomposes claims into atomic facts, Qwen3-Embedding-8B retrieves evidence from a Pathway vector store, DeepSeek-R1 judges contradictions. Structured Python object passing preserves similarity scores and claim provenance across agents. 98.3% success rate on 60 test cases. IIT Kharagpur Hackathon | Llama 3.3, DeepSeek-R1, Qwen3, Pathway |
| BTC-Forecaster | Fork of TradingAgents repurposed for intraday BTC forecasting. Multi-agent pipeline: technical + news + sentiment analysts → bull/bear debate → trader → risk → final forecast. Emits 1h/4h price predictions with a self-scoring track record logged against realized prices. | LangGraph, Python, Yahoo Finance |
| AMD AIPL | Fine-tuned Qwen3-4B using SFT + GRPO for a 1v1 Q-agent vs A-agent tournament. Custom reward function; self-play loop where question and answer models iteratively improve each other. AMD Hackathon at IIT Bombay | Qwen3-4B, PyTorch, GRPO |
| Project | What it does | Stack |
|---|---|---|
| AI-Avtaar | End-to-end character pipeline: upload photos → automated LoRA training → SDXL image generation → virtual clothing try-on. Four isolated Python environments orchestrated through a single Streamlit UI. | Kohya_ss, FLUX.1 / A1111, CatVTON, Streamlit |
| AI Video Creator | Three-stage generation pipeline: Llama 3.1-8B writes a storyboard → FLUX.1-dev renders per-scene images → Wan2.2-T2V animates them. Gradio tabbed interface. Tested on AMD MI300X with ROCm. | Llama 3.1, FLUX.1-dev, Wan2.2, Gradio, ROCm |
| Gender Detection API | FastAPI endpoint using RetinaFace for face detection and CLIP for gender classification. Handles multiple faces, non-human images, and mismatches as distinct typed error responses. | FastAPI, CLIP, RetinaFace, PyTorch |
| Project | What it does | Stack |
|---|---|---|
| ResumeTeX | Browser-based LaTeX resume builder: manual form, AI import from PDF/DOCX (extract → parse → verify pipeline), AI tailoring to job descriptions. Deterministic anti-fabrication guard restores all original facts post-AI. Deployed on Render. | Next.js 14, FastAPI, PostgreSQL, Nebius LLM |
| Stock News Summarizer | Scrapes TradingView, Finviz, and Polygon; Gemini Pro selects top articles and writes <500-word summaries with 7-day "what changed today" diffs. SQLite history, daily refresh at 8 AM IST. Free-tier deployed on Render. | Flask, Gemini Pro, SQLite |
| Student Performance Analysis | CSV upload → K-Means / Agglomerative clustering → per-student performance dashboard with trend charts, subject breakdowns, class comparisons, and Excel export. | React, Flask, Scikit-learn, MUI |
| Project | What it does | Stack |
|---|---|---|
| interactive-preview-skill | A Claude Code skill that turns any React/Next.js codebase into an interactive, theme-matched "try it before you sign up" demo a guided product tour on mock data, auto-generated from your own UI. Leaks no backend. | React, Nextjs, Tailwindcss |
| Support Finder | MV3 Chrome extension with a four-layer deterministic pipeline: DOM scan → same-domain path probing → schema.org extraction → confidence scoring. Returns ranked support contacts with explanations. No AI; no fabrication. | TypeScript, React, MV3 |
| android-compose-design | Agent skill for Claude Code. Guides AI to produce distinctive Jetpack Compose UI intentional color, type hierarchy, shape language, and motion instead of Material 3 defaults. | Jetpack Compose, Kotlin |
| Project | What it does | Stack |
|---|---|---|
| Solana Price Analysis | OHLCV data (2021–2024), 44-column feature set via the ta library, ML price prediction model, and live Binance price dashboard. |
Python, Streamlit, Binance API |
| Rossmann Retail Analysis | 1M+ row sales dataset: cleaning, feature engineering, EDA on Snowflake. Quantified 81.5% sales uplift from promotional periods. | Snowflake, SQL |
| Time Series Forecasting | Four-framework side-by-side: LSTM on temperature data (TF), airline passengers (PyTorch), ACGL stock SARIMAX (Statsmodels), MSFT stock LSTM (Keras). Flask web interface for the SARIMAX model. | TF, PyTorch, Statsmodels, Keras |
