|
||
ML Researcher at Arizona State University, architecting a modular CCP plasma-simulation framework (PyTorch Lightning + JAX) with swappable architectures, samplers, and collocation strategies for Applied Materials semiconductor R&D.
- Author of MACE-PINNs (M.S. thesis), a multi-network architecture for coupled-equation Physics-Informed Neural Networks.
- Previously built protein-thermostability & developability ML pipelines at OpenEye / Cadence Design Systems (ESM2 650M pLM, RAPIDS, contrastive learning at 50K+ batch scale).
- Shipped production GenAI / RAG systems at Talin Labs (fine-tuned Mistral-7B on Kubernetes, multi-agent LangChain, p95 < 200ms @ 10K users).
- M.S. Data Science (High-Performance Computing) @ ASU · GPA 3.72/4.0 · published in IEEE Access.
- Interests: Physics-Informed Neural Networks, LLM orchestration & agents, GPU-accelerated ML, edge/backend system design.
| Languages |
|
| AI / ML & GenAI |
-7400B8?style=for-the-badge&logo=nvidia&logoColor=white){kind=icon}
|
| Data & Backends |
|
| Cloud & DevOps |
|
- Architecting a modular CCP plasma-simulation framework with swappable architectures, samplers, collocation strategies, and interpolators for Applied Materials semiconductor R&D.
- Ran 60+ experiment configurations with adaptive loss balancing, identifying the optimal training setup through reproducible experiment tracking.
- Built an end-to-end protein-thermostability pipeline from scratch, 7.2 ms/seq across 1M+ sequences using ESM2 (650M-param pLM), Hugging Face, cuDF, and PyTorch Lightning.
- Developed a contrastive-learning architecture (cuML + RAFT replacing GPR), scaling batches 333× (150 → 50K+) on 5120-dim embeddings.
- Built a unified OmegaConf + Pydantic config framework parallelising 20+ antibody-developability experiments.
- Deployed fine-tuned Mistral-7B-Q8 on K8s across 12+ enterprise on-prem environments at p95 < 200ms for 10K users.
- Built a RAG evaluation framework over 10K human-evaluated queries reaching 88% accuracy (chunk precision, citation accuracy, cross-document consistency).
- Architected a 6-agent LangChain system with intent-based routing over FAISS + PDF/XLSX/DOCX parsing, cutting manual document review from weeks to minutes.
| Project | Description | Stack |
|---|---|---|
| Samhita ↗ · Knowledge-Backend Pipeline | PDF→knowledge-base pipeline turning 5,941 textbook pages into a 72K-node / 130K-edge graph with 53K BioLORD embeddings; ~100% figure/table extraction, Claude-Haiku enrichment via Anthropic Batches API. | Python Pydantic v2 Claude BioLORD |
| HybridFlow ↗ · Hybrid Retrieval RAG | Hybrid-retrieval RAG backend over a 93K-node Neo4j graph + Qdrant vectors; success@5 0.90, streaming Haiku→Sonnet pipeline at 14.7× throughput, 8-gate quality suite. | FastAPI Qdrant Neo4j Anthropic |
| sushrutalgs-bff ↗ · Edge BFF | 33 KiB Cloudflare Worker fronting iOS + web; edge JWT auth at p95 ~0.13ms, atomic plan-aware quotas via Supabase RPC, fail-closed under load. | TypeScript Hono Cloudflare Workers |
| sushrutalgs-ios ↗ · Native iOS Client | iOS 26 SwiftUI RAG chat client (80 views, Swift 6 strict concurrency); SSE typewriter streaming, cross-device handoff, 3 auth flows; 20.8 MB install. | Swift 6 SwiftUI Supabase |
| Yelp Recommendation Platform ↗ | PySpark ETL over the full 6.99M-review dataset at ~460K rows/sec; Spark ALS recommender + sentiment classifier; ~23,000× inference speedup via NumPy export. | PySpark FastAPI MLflow Docker |
| Title | Journal | Description |
|---|---|---|
| MACE-PINNs: A Multi-network Architecture for Coupled-Equations PINNs ↗ | M.S. Thesis, ASU (2025) | Parallel subnetworks with iterative residual constraints, Fourier-feature embeddings, and adaptive gradient-norm weighting; validated on Gray-Scott & Ginzburg-Landau 2D systems. |
| Classification of Potentially Hazardous Asteroids Using Supervised Quantum ML ↗ | IEEE Access, vol. 11 (2023) | VQC + PegasosQSVC at 98.11% accuracy / 92.69% F1 on 958K records. |
| MetaHate: AI-Based Hate-Speech Detection for Secured Online Gaming in the Metaverse ↗ | Security and Privacy, Wiley (2023) | Gradient boosting at 86.01% on a Hindi-English code-mixed dataset. |
ML Researcher @ ASU · ex-Cadence/OpenEye, Talin Labs · Physics-Informed ML, GenAI & High-Performance Computing