Focused on designing reliable AI workflows using retrieval, tool-use, and structured reasoning.
- LLM Alignment & Training: Engineering robust pipelines for classic fine-tuning and state-of-the-art reinforcement learning from human/ai feedback (RLHF/RLAIF).
- Agentic AI Workflows: Building production-ready, tool-augmented systems utilizing self-correction, tool-use, and programmatic validation to neutralize LLM hallucinations.
- Multimodal Architectures: Designing unified, end-to-end pipelines that orchestrate text, vision, audio, and video synthesis for complex industry verticals.
-
🚀 transformer-architect A unified, production-grade LLM training and alignment framework. Features modular implementations for downstream tasks (BERT), Direct Preference Optimization (DPO), and Group Relative Policy Optimization (GRPO reinforcement learning pipelines) to streamline model alignment and optimize memory overhead.
-
🧮 Tool-Augmented-AI-Mathematical-Reasoning-System A tool-augmented LLM reasoning system designed to eliminate hallucinations in multi-step complex problem-solving. Features an advanced orchestrator for multi-step prompt decomposition, LaTeX parsing, and a secure programmatic validation loop using Python execution sandboxes.
-
🤖 WhyBuddy [The Gemma 4 Good Hackathon] An end-to-end multimodal educational AI assistant powered by Gemma 4. It orchestrates an advanced, unified pipeline combining natural language reasoning, vector-based knowledge retrieval (RAG), image generation, text-to-speech (TTS), and video synthesis for real-time interactive learning.
-
💼 ElevioCareer A production-grade, AI-driven enterprise web application that leverages LLMs to parse PDF resumes for precise, embedding-based ATS compatibility scoring, while utilizing asynchronous web scraping to dynamically recommend the top 5 most relevant job listings.
-
🧬 RNA 3D Structure Prediction - Kaggle Built an end-to-end deep learning pipeline utilizing Protenix-based architectures for molecular structure optimization. Implemented rigorous molecular sequence preprocessing, heavy feature engineering, and ensemble-based refinement during the competition validation phase.
-
🌌 Seti Image Classification Developed an end-to-end Convolutional Neural Network (CNN) classification and inference pipeline optimized for detecting anomalies in Seti signal spectrogram images.
