Projects
Research systems and engineering projects — verified work, with code where public.
CAD Code Generation with a Frozen LLM (CORE Robotics Lab) ↗
Closed-loop agentic pipeline for parametric CAD: a frozen LLM writes CadQuery, an execution environment compares the geometry against the ground-truth part, and a bounded LangGraph repair loop (up to 3 turns) corrects errors. Mean voxel IoU rises from 0.82 to 0.92 (expert prompts) and 0.57 to 0.70 (beginner prompts) over single-shot generation, n=200 per group, and +0.05 over a self-debug baseline. Repair is ground-truth-assisted, and the negative results (feedback-matrix tie, failed pre-registered gate) are reported as negatives. Ongoing.
Stack: Python, LangGraph, CadQuery, Pydantic, pytest
CERBERUS — Vision-Language Alignment & QA Fine-tuning ↗
Vision-text alignment with frozen CLIP ViT-L/14 and MPNet encoders, Matryoshka embeddings and InfoNCE: 77.7% / 77.5% image-to-text / text-to-image R@5 at 4096-d on PixMo validation, 62.9% / 64.0% at 128-d. LoRA SFT (r=64) of Qwen-2.5 7B for image QA, with training loss falling from 0.742 to 0.103; no held-out QA metric yet. A Perceiver Resampler ablation is reported as a negative result. Ongoing.
Stack: PyTorch, PEFT/LoRA, CLIP, SBERT, Matryoshka
ARTEMIS — VLM Routing Benchmark ↗
Offline research on routing requests across vision-language models. Built a leakage-corrected benchmark of 67.7K questions and 338K responses over a five-model pool (3,909 cross-split image hashes voided earlier numbers). On the test set an oracle router reaches 65.4% versus 49.1% for the best single model, but the learned router (50.9%) did not beat a per-task lookup (56.4%) — reported as a negative result.
Stack: PyTorch, Python, PostgreSQL, Weights & Biases
ATHENA — Storyboard Generation (team) ↗
Team project at Georgia Tech. I authored Production V1: a 14-node deterministic LangGraph pipeline with routing from an execution plan, Pydantic contracts, JSON persistence, and FastAPI/SSE, NiceGUI and CLI front ends (~800 tests collected). Release candidate with known open issues: the evaluator is text-only and failed image calls need re-verification. Manuscript in preparation.
Stack: Python, LangGraph, Pydantic, FastAPI, NiceGUI
SHASTRA — Agent Trace Modeling (team) ↗
Team research project (CS 8903, Spring 2026). I built the analysis stack that turns agent traces into blocks, LLM-labelled capabilities with an append-only label registry, and task-dependency graphs. Early stage: no workflow-reuse findings yet, and retrieval/adaptation is not implemented.
Stack: Python, LangGraph, Pydantic
Malware Analysis: Memory Forensics with ML ↗
Memory-forensics and ML research for ransomware detection in IoT-enabled energy systems, with Sonika Malik. Volatility 2 feature extraction with vendored YARA rules over 976 memory windows; nine scikit-learn classifiers compared, best Gradient Boosting at 88.74% accuracy (exploratory, with a leakage caveat). Paper accepted at IEEE ICAIA 2026. Dataset hosted on Hugging Face (33 GB analysis, ~470 GB raw memory dumps).
Stack: Python, Volatility, YARA, scikit-learn, pandas, Hugging Face Hub
Neural-Malware-Shield (team of 4) ↗
Graph-based malware classification on MalNet-tiny. I owned the GNN work: GraphSAGE reaches 0.781 accuracy and 0.777 macro-F1 (GAT: 0.708). Coursework team project.
Stack: Python, PyTorch Geometric
RL Soccer ↗
Academic team project: PPO curriculum training over 30M steps for 2v2 soccer on Ray/RLlib, evaluated over 10 episodes (8–2 against the baseline, 6–4 against an earlier curriculum). Small-N results, coursework (CS 8803 Deep RL).
Stack: Python, Ray/RLlib, PPO, DQN, Unity ML-Agents
AI-Security (CS 8803 ML Security) ↗
Coursework: PGD adversarial attacks (clean accuracy ~92% falling to 8.5% at 8/255), embedding poisoning (99.67% poison success, 93.30% clean accuracy), and membership inference on Pythia-160m (AUC 0.994). Blind-backdoor notebooks wrap the upstream backdoors101 code and are not an original implementation.
Stack: Python, PyTorch, Jupyter
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