Projects
Research systems and engineering projects — verified work, with code where public.
CAD Code Generation Pipeline (CORE Robotics Lab) ↗
Closed-loop agentic pipeline for parametric CAD code generation: an LLM generates CadQuery programs, an execution environment returns compile and geometry feedback, and a LangGraph-orchestrated repair loop iteratively corrects errors — inference-time search with verifiable geometric reward (frozen model, no fine-tuning). Includes geometric verification (Chamfer/Hausdorff on STL meshes), VLM-as-judge visual judgment, RAG-augmented prompts, and a pytest evaluation harness.
Stack: Python, LangGraph, CadQuery, Pydantic, VLMs, RAG, pytest
Malware Analysis: Memory Forensics with ML & Volatility ↗
Research project integrating machine learning with memory forensics (Volatility Framework) and YARA rules for ransomware detection in IoT-enabled energy systems. Automated analysis of 18 memory dumps (6 benign, 12 ransomware families) with 100+ YARA rules. Published at IEEE ICAIA 2026. Dataset hosted on Hugging Face (33 GB analysis + ~470 GB raw memory dumps).
Stack: Python, Volatility3, YARA, scikit-learn, pandas, Jupyter, Docker, Hugging Face Hub
ARTEMIS — Cost-Aware VLM Routing (Which-VLM-Router) ↗
Neural multi-task router that dynamically selects the optimal Vision-Language Model per request, with SLA-aware load balancing (simulation-validated). Serves 5 VLMs across 10 vLLM endpoints over VQA, OCR, captioning, and reasoning tasks with 5 dynamic routing modes; ~340K routing profiles across ~68K queries, with 90.3% oracle-utility recovery in balanced mode.
Stack: PyTorch, vLLM, FastAPI, PostgreSQL, Weights & Biases
CERBERUS — Vision-Language Alignment for Edge (public repo: CEREBRUS) ↗
Cross-modal alignment with frozen encoders (CLIP ViT-L/14, SBERT) using Matryoshka Representation Learning compressing embeddings 4096→128 dims (~78% R@5 at 4096-d on PixMo retrieval). LoRA fine-tuning of Qwen2.5 decoders; PyTorch DDP distributed training with mixed precision on H100/A100 clusters. Includes an honestly-reported negative Perceiver Resampler ablation.
Stack: PyTorch (DDP, mixed precision), PEFT/LoRA, CLIP, SBERT, Matryoshka
RL Soccer ↗
Academic project (team): PPO agents (baseline, reward shaping, curriculum learning, self-play) and a DQN baseline for 2v2 soccer using Ray/RLlib.
Stack: Python, Ray/RLlib, PPO, DQN, Unity ML-Agents
CAD-Physics — FEA-Guided Verification of LLM-Generated CAD ↗
Extends CAD verification from geometric validity to physical validity: finite-element-analysis checks (stress, displacement, safety factors) on LLM-generated designs. Early-stage research extension of the CORE Lab CAD pipeline.
Stack: Python, Jupyter, FEA tooling
AI-Security ↗
Security concepts in AI: adversarial attacks (PGD), embedding poisoning, blind backdoors, model extraction, membership inference on LLMs, jailbreaks, and watermarking.
Stack: Python, Jupyter Notebook
Audit_Script_Development ↗
Automated security audit script for Linux environments to identify vulnerabilities and assess security posture.
Stack: Shell, Python
Machine_Learning ↗
A collection of small projects and experiments undertaken for learning machine learning concepts.
Stack: Python, Jupyter Notebook