arXiv2026Under Review
Can Small Models Reason About Legal Documents? A Comparative Study
LLMsReasoningLegal AISmall Models
Forward Deployed AI Engineer at Oscar Health
AI for healthcare teams
5+ years across data engineering, ML, and AI
At Oscar Health, I build tools for code review, claims analytics, and provider contract extraction. Previously at ModMed, I built clinical AI for 15,000+ providers. My research covers LLM reasoning, hallucination detection, and computer vision.
Stuff I've Built
From self-driving cars to medical AI and agentic applications.
Here's what I've been building.
🌐 Web Apps2025
Resume2Portfolio
Transform your resume into a beautiful portfolio website in minutes. No coding required.
Next.jsAIAutomation+2+1
🧠 GenAI & LLMs2025
MEDHALT
Framework for evaluating AI-generated medical SOAP notes. Detects hallucinations and validates clinical accuracy.
PythonNERLLM+2+1
🤖 Agentic AI2026
AI Influencer Bot
AI-powered social media bot that auto-posts AI news from HackerNews. Autonomous content curation and posting.
PythonLLMAutomation+2+1
🧠 GenAI & LLMs2024
RAG-Anything
Ask any question to your documents, PDFs, or SQL databases. One RAG to rule them all.
PythonLangChainRAG+2+1
🌐 Web Apps2026
HackSwipe
Tinder-style app for discovering winning hackathon projects. Swipe through 220+ projects with YouTube demos.
JavaScriptWeb AppDevpost+2+1
🌐 Web Apps2026
H1B Wage Finder
Find lowest H1B prevailing wages for your occupation across US metros. Better wages = better approval!
PythonData AnalysisImmigration+2+1
Research & Publications
Published and under-review work across LLM reasoning, interpretability, and applied computer vision.
arXiv2026Under Review
Do Hallucination Neurons Generalize? Evidence from Cross-Domain Transfer in LLMs
LLMsHallucinationInterpretability
Plants2025PublishedFirst Author
Detecting Escherichia coli Contamination on Plant Leaf Surfaces Using UV-C Fluorescence Imaging and Deep Learning
Computer VisionYOLOv11AgricultureFood Safety
IEEE ICCMST2023Published
An Effective Model for Smartphone-Based Pothole Classification and Admin Alerting System
Computer VisionEdge AIMobile
IEEE2022Published
ECG-Based Early Heart Attack Prediction Using Neural Networks
Healthcare AINeural NetworksTime SeriesCardiology
My Toolkit
The technologies I use to turn caffeine into code
🔧
Agent Tools & Evaluation
MCP
Cube
Langfuse
LLM-as-Judge
Codex
OAuth2
Prometheus
PagerDuty
🤖
GenAI & LLMs
LangChain
OpenAI API
Claude
RAG Systems
Hugging Face
Fine-tuning
Prompt Engineering
Vector DBs
🧠
Machine Learning
TensorFlow
PyTorch
Keras
Scikit-learn
Computer Vision
NLP
Deep Learning
MLOps
⚡
Data Engineering
Python
SQL
PySpark
Kafka
Airflow
AWS (S3, Redshift)
Snowflake
dbt
💻
Development
Python
TypeScript
React/Next.js
FastAPI
Node.js
Swift (iOS)
Git
Docker
☁️
Cloud & Tools
AWS
GCP
Vercel
GitHub Actions
Jupyter
VS Code
Linux
Streamlit
Where I've Worked
AI tools at Oscar Health, clinical systems at ModMed, and earlier work in data engineering and research.
InternshipJun 2026 - Present
Forward Deployed AI Engineer
- ▹Built an MCP code review server covering 900+ pull requests. It flags server-side violations, missing auth scopes, and prompt-injection risks, with Langfuse traces on every run. Reduced review time by 30% and caught 40+ compliance issues on its first day.
- ▹Built a Cube semantic layer over dbt, exposed through MCP with OAuth2 scopes, for natural-language queries over governed claims metrics. Evaluated answers against a 500-query golden set to catch metric drift.
- ▹Built multimodal extraction for 14,000+ provider contracts. Agents re-check source pages when validation fails or answers conflict, improving first-pass accuracy from 82% to 95%. A labeled evaluation set gates model and prompt changes.
- ▹Delivered plugins, MCP servers, and skills for legal, finance, and people teams, and rolled out Codex to engineering with admin policies and per-user usage tracking. Added Prometheus metrics, Langfuse traces, and PagerDuty escalation, bringing detection of degraded tool calls below 2 minutes.
MCPCubedbtOAuth2LangfuseLLM-as-JudgeCodexPrometheusPagerDuty
Full-timeFeb 2025 - Apr 2026
AI Engineer, GenAI Applications & LLM Systems
- ▹Shipped Clinical Ambient AI Scribe serving 15,000+ providers across 11 specialties (400K+ daily encounters), automating 70% of documentation via real-time transcription + LLM SOAP generation
- ▹Built agentic document pipeline (OpenAI Agents SDK + fine-tuned Qwen2-VL VLM) routing 10M+ clinical pages/month, replacing a $400K/month vendor with a $20K in-house system (95% cost reduction)
- ▹Built Text2SQL + clinical knowledge graph over ModMed’s EHR warehouse (200+ tables, pgvector embeddings), cutting analyst request volume by 60%
- ▹Architected production multi-agent RAG with LoRA-finetuned SLMs, hybrid pgvector + BM25 retrieval, and cross-encoder reranking, achieving 94% retrieval precision across 50K+ clinical documents
- ▹Open-sourced MEDHALT for clinical hallucination detection (DeBERTa NER + LLM-as-judge) achieving 92% accuracy vs. GPT-4, with MLflow tracking and golden-set regression testing
- ▹Shipped LangChain/LangGraph/Claude monitoring framework for Scribe quality, cutting incident response to under 5 minutes with PHI-safe pipelines and prompt-injection filtering
PythonLangChainLangGraphVLMsRAGpgvectorDatabricksKubernetesMLflow
InternshipMay 2024 - Jul 2024
AI Software Developer Intern
GeoSpider AI•USA (Remote)
- ▹Built LangGraph multi-agent RAG that autonomously resolved 65% of customer tickets across a 50K-doc knowledge base, with dynamic routing and LLM-as-judge scoring
- ▹Built LLM-as-judge routing layer with dynamic few-shot prompting, improving helpfulness from 43% → 76% and relevance by 30%
- ▹Designed FAISS + keyword hybrid search with semantic reranking, achieving 92% recall@10, serving 150+ concurrent users via vLLM with 40% p95 latency reduction
- ▹Implemented Redis-backed multi-turn agent memory and FastAPI inference gateway with fallback logic and structured logging
LangGraphFAISSvLLMRedisFastAPILLM-as-Judge
ResearchFeb 2023 - Dec 2024
Graduate Researcher, AI/ML
- ▹Developed hybrid YOLOv8-ViT model improving small-object detection by 15% with Grad-CAM / EigenCAM explainability. Published at SPIE 2025 and IEEE 2023; presented at both
- ▹Built React dashboard with Grad-CAM visualizations replacing static PDF reports (adoption 15% → 85%). Prototyped CLIP-based multi-modal retrieval between lab images and research reports
- ▹Automated model retraining via MLflow + GitHub Actions CI/CD, cutting deployment from 4 hours to 15 minutes
PythonYOLOv8ViTCLIPMLflowGrad-CAM
Full-timeJun 2021 - Dec 2022
Software Data Engineer
- ▹Engineered BERT + XGBoost intent classifier (88% F1) predicting technician dispatch necessity. Eliminated 12K unnecessary dispatches/year, saving $2M annually
- ▹Built Elasticsearch + Word2Vec anomaly detection for network telemetry, cutting diagnosis time by 40% and Tier-2 escalations by 25%
- ▹Optimized PySpark / Delta Lake pipelines processing 1M+ logs/day (30% latency reduction); designed Azure Synapse warehouse with dbt
- ▹Built NLP ticket-categorization pipeline with fine-tuned BERT, achieving 91% accuracy, 500K+ monthly interactions
PythonPySparkBERTAzure SynapseDelta LakeElasticsearch
Education
Master of Science in Computer & Information Science
University of Florida
Gainesville, FL2023 - 2024
Focus: Machine Learning, Computer Vision, Data Engineering
Bachelor of Technology in Computer Science
GPA: 3.9/4.0GITAM University
Visakhapatnam, India2017 - 2021
Focus: Software Engineering, Data Science
Get in touch
Got an interesting project?
Let's talk!
Reach out to discuss healthcare AI, model evaluation, research, or a project you are working on.
Drop me an email