AI/ML Engineer for LLMs & Agentic Systems, Autonomous Driving, and Production ML.

Agentic platform for non-technical users to explore company financials, stock prices & trends via text + voice, delivering insights with charts, tables, and cited summaries.

• • LangChain + LangGraph tool orchestration, SQL generation, web search fallback, guardrails & transparent citations.
• • Realtime APIs + voice workflow (Whisper) for conversational finance UX.

Retrieval-Augmented Generation system integrating BioGPT with NCBI Entrez for biomedical question answering; ~78% accuracy on PubMedQA.

• • Built retrieval + generation pipeline, evaluation workflow, and interactive UI demo.
• • Reported peak batch accuracy during evaluation and compared with baselines.

Implemented VectorNet (two-stage GNN) for motion prediction using Argoverse 1.0, including polyline SubGraphs and global self-attention.

• • Hand-crafted features from raw CSVs; built per-scenario graph datasets for ego, lanes, and surrounding agents.
• • Visualized and evaluated predicted trajectories vs. ground truth.

Two-stage evaluation pipeline profiling open-source LLMs on toxicity, bias, truthfulness, and safety; plus an advisor that recommends models based on user priorities.

• • Reproducible trust profiles for ethical deployment; explained trade-offs across safety/fairness/factual reliability.
• • Built Streamlit dashboard with plots and an LLM-powered explanation layer.

Trained Faster R-CNN and YOLOv8 on Waymo and mixed Waymo+KITTI datasets with unified 3-class taxonomy; improved cross-dataset consistency and mAP.

• • Faster R-CNN: tighter boxes + better occlusion handling.
• • YOLOv8: ~3× faster inference with strong accuracy.

Integrated Graph-of-Thought reasoning into the LightEMMA VLM trajectory prediction framework, adding multi-intent generation, branch scoring, and refinement to improve motion forecasting on nuScenes vs CoT baselines.

• • Used Claude Sonnet 4.5 and Google Gemini Flash 2.0 for reasoning + scoring workflow.
• • Focus: higher-quality forecasting via structured branching and refinement.

Three-tier Hierarchical RL system for multi-asset portfolio optimization using market data + Reddit sentiment to improve risk-adjusted returns and reduce drawdowns over 3 years of backtesting.

• • Trained DQN, PPO, DDPG, TD3 agents and aggregated via meta- and super-agent layers.
• • Integrated FinBERT sentiment with Yahoo Finance pipelines; beat equal-weighted and single-agent baselines.

Built a portable real-time AI device to classify benign vs malignant skin cancers using MobileNetV2; achieved 94.2% accuracy on the final prototype using Intel® Movidius™ NCS + Raspberry Pi.

• • Fine-tuned MobileNetV2 and deployed an edge inference pipeline for low-resource healthcare settings.
• • Full prototype: camera module + on-device inference + optimized graph deployment.