AI Engineer • LLM Fine-Tuning • AI Agents & Evaluation
Building and evaluating intelligent systems that learn, reason, and act.
AI engineer and final-year B.Tech student with hands-on experience in LLM fine-tuning and reinforcement learning, AI agents, and model evaluation. Currently creating and reviewing AI benchmark tasks at Ambiguity Labs, while building practical AI tools for browser automation, web editing, and data analysis.
B.Tech, AI & Machine Learning
LD College of Engineering · Ahmedabad
Core technologies and frameworks I work with
A selection of AI & ML projects I've built
Fine-tuned Qwen2.5-0.5B for web navigation using SFT followed by GRPO, raising accuracy from a 25% zero-shot baseline to 100% on the project's BrowserGym test suite. Built the Docker training setup and custom reward functions to run on a single 4 GB VRAM GPU.
Built a visual web editor that turns plain-English UI requests into live page updates using Gemini Flash or local models through Ollama. Created a CSS patching system that changes only the requested styles without breaking the page layout or interactive elements.
Built a multi-agent analytics tool that takes a plain-English question and handles data queries, transformations, and chart generation. Added local model support through Ollama so sensitive data can be analyzed offline.
Deep learning framework for manipulated media detection using domain-adversarial training, improving cross-domain accuracy by 15% on unseen data distributions. Curated 10,000+ sample dataset spanning FaceSwap, Face2Face, and NeuralTextures generation techniques for robust evaluation.
Professional journey in AI & Machine Learning
Remote
Ahmedabad
Academic foundation in AI & Machine Learning
L.D. College of Engineering, Ahmedabad
Relevant Coursework
Key milestones and accomplishments
Improved Qwen2.5-0.5B from a 25% zero-shot baseline to 100% accuracy on the project's BrowserGym test suite.
Creates and reviews measurable benchmark tasks for evaluating AI systems at Ambiguity Labs.
Built a Dockerized SFT and GRPO training setup that runs on a single 4 GB VRAM GPU.