MS Computer Science @ Indiana University, Bloomington
Research Assistant — Cybersecurity AI @ Indiana University, Kelley School of Business
- Building a hierarchical LLM + RL red team adversary for autonomous penetration testing in the CybORG CAGE Challenge 4 enterprise network (9 subnets, 1 red vs 5 blue defenders + 48 green agents)
- Designed a 3-layer reward shaping system (negated opponent rewards + milestone bonuses + prerequisite penalties) to enable RL training from zero environment signal
- Evaluated 13 LLMs (0.6B–70B) across 4 inference backends on NVIDIA H100 GPUs - found that no LLM achieves meaningful attack success zero-shot, and model size does not correlate with performance
- PPO with action masking | vLLM | Curriculum learning | SLURM/HPC on Big Red 200
- Presenting at SPIE Defense + Security 2026 - April 2026, National Harbor, MD
|
Agentic Cybersec Threat Analyst
3-agent LangGraph pipeline that transforms raw CVEs into defense playbooks - ingests NVD, CISA KEV & OTX feeds, maps threats via hybrid RAG over 19K+ MITRE ATT&CK techniques, and auto-generates Sigma detection rules. FastAPI + React + Qdrant.
|
FinSight
Four specialized LangGraph agents (Document, Quantitative, Risk, Synthesis) dissect SEC filings (10-K, 10-Q, 8-K) with hierarchical PageIndex navigation. Local Ollama + Groq cloud fallback. FRED & Finnhub market data. 129 unit + 10 E2E tests.
|
|
FinSent-CoT
16,944-sample balanced dataset pairing financial texts with expert Chain-of-Thought reasoning distilled from Qwen3-235B on H100 GPUs. SFT + GRPO formats for on-device LLMs (0.5B–8B). Multi-stage QA with 15.5% rejection rate.
|
FinistralAI
Mistral-7B fine-tuned for financial sentiment with LoRA adapters (rank 16, alpha 32). BF16 precision, gradient checkpointing & DeepSpeed multi-GPU training. Published on HuggingFace Hub.
|
|
DDoS Mitigation
Three-tier defense-in-depth: XDP/eBPF kernel-level rate limiting (10M pps), P4 BMv2 in-network per-flow detection (1024 flows), and BGP FlowSpec/RTBH upstream blackholing. Tested on Jetstream2 with Mininet + FRR.
|
Brain Tumor Classification ★ 8
CNN classifying brain tumor MRI scans into 4 categories (Glioma, Meningioma, Pituitary, No Tumor) across 7K+ images. End-to-end pipeline from Kaggle download to prediction.
|
|
Drug Typology Classification ★ 8
Comparative ML study - Logistic Regression, Random Forest, SVM & Voting Ensemble with precision-recall trade-off analysis for minimizing clinical false positives.
|
FinEdu.AI ★ 13
Conversational financial education assistant powered by LLaMA-2-7B with RAG over curated financial terminology. Gradio web UI with beginner-friendly explanations and real-world examples.
|
| Research Assistant - Cybersecurity AI @ Indiana University, Kelley School of Business |
Hierarchical LLM+RL red team adversary for CybORG CAGE Challenge 4 - RL controller hit 8x heuristic baseline via custom 3-layer reward shaping; benchmarked 13 LLMs (0.6B–70B) across 4 backends on H100s; Presenting @ SPIE 2026 paper |
| Research Assistant - Software Developer @ Indiana University, Kelley School of Business |
LLM annotation pipeline with Tree-of-Thoughts + Self-Consistency - labeled 167K+ patent abstracts & 22K+ AV disengagement incidents with 7-way classification using OpenAI & Gemini APIs |
| Software Developer - AI @ OCG Technologies, Singapore |
Fine-tuned Llama 2 chatbot with LoRA/PEFT on SageMaker - 83% accuracy boost; RAG with OpenSearch for 75% support overhead reduction |
| Software Developer - Data @ Visual Labs, Mumbai |
3 TB drone imagery ETL pipeline — PySpark + OpenCV preprocessing, AWS Lambda orchestration, DynamoDB storage, Power BI dashboards; 45% stakeholder outcome improvement |
| Languages | Python Java C SQL JavaScript P4 HTML/CSS |
| ML / DL | PyTorch TensorFlow Scikit-learn CNNs XGBoost |
| RL | PPO GRPO GSPO ORPO Curriculum Learning |
| LLMs | vLLM SFT PEFT LoRA/QLoRA RAG LangGraph Transformers Ollama llama.cpp OpenAI API Gemini API |
| Infra | AWS SageMaker Lambda S3 DynamoDB OpenSearch SLURM/HPC (H100) Docker |
| Web / APIs | FastAPI React Streamlit Gradio |
| Data | PySpark Pandas NumPy OpenCV BeautifulSoup |
| Visualization | Power BI Matplotlib Seaborn Altair |
| Tools | Git Weights & Biases DeepSpeed pytest Unsloth HuggingFace Hub |