Reinforcement Learning · Sustainable AI · LLM Systems · Retrieval-Augmented Reasoning
🎓 PhD IIIT-D | 🏛️ University of South Dakota AI Research Lab | 📜 Google Scholar | 💼 LinkedIn
♻️ AI-CARE — Carbon-Aware ML Evaluation Benchmark
The first systematic benchmark for measuring the carbon footprint of machine learning model evaluation.
AI-CARE introduces two novel sustainability metrics — SCAS (Sustainable Carbon Accuracy Score) and CATC (Carbon-Aware Training Cost) — enabling rigorous comparison of accuracy vs. environmental impact across diverse architectures and datasets.
| Dimension | Details |
|---|---|
| Architectures | 6 (CNNs, Transformers, hybrid models) |
| Datasets | 5 benchmark datasets |
| Key Metrics | SCAS · CATC · Carbon Efficiency |
| Focus | Sustainability · Green AI · Reproducibility |
🤖 AI-CARE LLM — LLM Billing Mismatch Detection Pipeline
An automated pipeline detecting billing discrepancies in LLM API usage via the NRP Nautilus HPC cluster.
This project extends the AI-CARE framework to large language models, exposing token-billing mismatches across commercial and open-source LLM providers by running inference at scale on Nautilus infrastructure.
| Dimension | Details |
|---|---|
| Models | Qwen3 · GPT-OSS · MiniMax-M2 |
| Infrastructure | NRP Nautilus (HPC/GPU cluster) |
| Task | Billing mismatch detection · Token auditing |
| Focus | Transparency · Cost Accountability · Open Science |
🔍 Reflexion / RAR — Retrieval-Augmented Reflexion
Extends the Reflexion self-refinement framework with retrieval-augmented reasoning, evaluated on HotPotQA, ALFWorld, and HumanEval Hard.
RAR (Retrieval-Augmented Reflexion) augments the agent's verbal reinforcement loop with dynamically retrieved context, improving multi-hop reasoning and code generation over vanilla Reflexion baselines.
| Benchmark | Task Type | Improvement over Baseline |
|---|---|---|
| HotPotQA | Multi-hop QA | ✅ Retrieval-boosted accuracy |
| ALFWorld | Embodied planning | ✅ Higher task-completion rate |
| HumanEval Hard | Code generation | ✅ Pass@k on hard problems |
- Postdoctoral Researcher at the University of South Dakota (USD) AI Research Lab.
- Research interests: Reinforcement Learning, Sustainable AI, World Models, LLM Systems, and ML in Healthcare.
- Author of the upcoming textbook: Reinforcement Learning Fundamentals: From Theory to Practice (+ companion code repo).
- Organizer of the AI Symposium @ USD and passionate about teaching, mentoring, and community building.
- ✍️ Completing a comprehensive RL textbook (LaTeX source + reproducible code)
- 🌱 Exploring world models and sample-efficient embodied RL (DreamerV3, AdaWorld)
- 📊 Collaborating on AI for biomedical computation at USD
- 🧪 Extending AI-CARE to edge inference and federated learning settings
| Repo | Description |
|---|---|
| Reinforcement-Learning-Explained-Code | 📚 Companion code for my RL textbook |
| AI-Symposium | 🎤 Website for USD AI Symposium |
Languages: Python · LaTeX · SQL
Libraries: PyTorch · TensorFlow · scikit-learn · HuggingFace Transformers
Infrastructure: HPC (Lawrence @ USD) · NRP Nautilus · Git · Overleaf
I teach meditation 🧘 alongside AI research — cultivating clarity of mind and clarity of models, one step at a time.
⭐ Check out the pinned repositories above for active research!