Hello! I'm
SUR
VAGHASIYA
An
AI/ML Engineer
Data Scientist
About Me
I like problems that begin with, “There has to be a smarter way to do this.” That curiosity has taken me from forecasting and predictive analytics to RAG systems, voice AI, real-time platforms, cloud infrastructure, and full-stack products.
I'm pursuing Data Science at Stony Brook University, but I rarely stay inside one box. I enjoy moving between data, models, APIs, systems, and interfaces until an idea becomes something practical, reliable, and genuinely useful.
I've worked across machine learning, generative AI, backend engineering, MLOps, geospatial analytics, and Web3 — but I care less about listing technologies and more about what they make possible.
I'm drawn to ambitious problems, thoughtful teams, and work that sits at the intersection of AI, data, and software.
WHAT
I DO
AI & DATA SCIENCE
ML pipelines, forecasting & RAG systems
End-to-end ML — ensemble forecasting (Prophet + XGBoost + ARIMA) in Insight-X, multimodal RAG with FAISS in EchoMindAI, and geospatial ML with SHAP explainability on 33K records.
Skillset & tools
VOICE AI & AGENTS
Real-time speech & autonomous agents
SAARTHI (HackRare 2026, Top 4) — clinical voice AI with Retell AI + Claude. EchoMindAI — 16 agentic tools, 7.5× STT latency cut via Groq LPU, built on LangChain ReAct + FastMCP.
Skillset & tools
FULL-STACK & MLOPS
APIs, web apps & cloud deployment
Shipped full-stack products — Next.js + FastAPI + Docker stacks, AWS CI/CD, and a Web3 escrow on Ethereum Sepolia with Solidity + Hono.js + Cloudflare Workers.
Skillset & tools
My career &
experience
Data Science & AI Intern
Adani Enterprise Ltd.
2025
Jan – Jul 2025 · Ahmedabad, India
Shipped a production AI forecasting engine — 78% accuracy on sales & revenue prediction using Python, Scikit-learn, and AWS. Automated the full ML lifecycle (retraining → EC2/S3/Lambda deployment), cutting operational overhead by 37%. Built dashboards that replaced manual reporting for 10+ product teams, saving 34% in analysis time.
Python Machine Learning Intern
Techmicra Data Systems
2024
Jun – Jul 2024 · Ahmedabad, India
Built ML pipelines from scratch — data cleaning, EDA (ANOVA, correlation matrices), and regression modeling (Linear, Ridge, Polynomial). Grid Search CV cut overfitting by 35% and improved generalization by 27%. LASSO/Ridge regularization pushed model interpretability up by 42%.
Web Developer Intern
Shaligram Infotech
2023
Jun – Jul 2023 · Ahmedabad, India
Built and shipped 5 responsive cross-browser web pages with 48% faster load times. Delivered 2 MySQL-backed WordPress sites for real clients — improving delivery speed and client satisfaction by 45%.
Education
MS, Data Science
Stony Brook University · New York
2027
Aug 2025 – May 2027 (Expected) Coursework: ML, Statistical Computing, Probability & Statistics, Data Analysis & Management, CS Fundamentals. Led research on LI Traffic Accident Analytics — 33K+ records, DBSCAN, XGBoost (AUC-ROC 0.72), SHAP, GeoPandas, NetworkX.
BTech, Computer Science & Engineering
Indus University · Ahmedabad, India
2025
Sep 2021 – May 2025 · GPA 3.4/4.0 Coursework: ML, NLP, Deep Learning, DSA, DBMS, Cloud Computing, Web Technology, Software Engineering. Covered the full CS spectrum — from algorithms & systems to ML, cloud, and applied engineering.
Honors & Awards
2026
Built SAARTHI — AI voice clinical monitoring platform for rare-disease symptom tracking. Combined Retell AI phone calls, Claude transcript analysis, FastAPI/Flask, PostgreSQL, personalized ML models, and a Next.js doctor dashboard.
My Work
01
LI Traffic Analytics
Data Science12 crash hotspot clusters (DBSCAN + Haversine on GPS) and 4 mixed-feature risk archetypes (FAMD + k-Means) from 33,228 Long Island accident records. XGBoost severity classifier: AUC-ROC 0.7185, 71% accuracy, 5-fold CV 0.7164 ± 0.0042. SHAP attribution + GeoPandas/osmnx road-network overlays for logistics risk scoring and safer routing.
Tools and features
Python · R · XGBoost · DBSCAN · SHAP · GeoPandas · osmnx · NetworkX · scikit-learn
02
SAARTHI
AI / VoiceVoice-first clinical monitoring for ENS rare-disease symptom tracking. Retell AI automates daily patient calls; Claude extracts 7 validated clinical scores with symptom-minimization and temporal-context detection. Tri-level ML: Global RF → personalized model (30+ readings) → Welford Z-score override. Deployed on Vercel + Railway.
Tools and features
Python · FastAPI · Flask · Claude AI · Retell AI · Next.js · FastMCP · PostgreSQL · scikit-learn
03
InsightX
ML / Analytics33% accuracy gain via Prophet + XGBoost + ARIMA/STL ensemble with P10/P90 confidence intervals, KS drift detection, and rolling MAE/RMSE backtests. Redis Pub/Sub + WebSocket eliminates HTTP timeouts on 5–30s ML jobs. War Games simulator stress-tests 4 economic scenarios in real time. A2UI/MCP AI copilot renders React charts inside chat. 7-service Docker stack.
Tools and features
Prophet · XGBoost · ARIMA · Redis · FastAPI · Next.js · Docker · MLflow · MinIO
04
EchoMindAI
Multimodal RAGMultimodal RAG with 16 agentic tools — DALL-E 3 vision, live stocks, news, 100+ language translation, Python REPL, maps. Groq LPU cut STT latency 7.5× (1500ms → 200ms). FAISS IndexFlatL2 + HuggingFace Sentence Transformers over PDF/DOCX/CSV/PPTX. 4-layer heuristic recovery system reduced rendering failures 90%.
Tools and features
LangChain · GPT-4o · FAISS · Groq · DALL-E 3 · OpenAI TTS · Streamlit · HuggingFace
05
Decentralized Escrow
Blockchain / Web3Full-stack Web3 escrow on Ethereum Sepolia. Solidity locks ETH at creation — dual-party approval releases funds, refund logic prevents permanently stuck transactions. Serverless API via Hono.js + Cloudflare Workers + Upstash Redis. React + ethers.js + MetaMask frontend; Hardhat validates all contract state transitions.
Tools and features
Solidity · ethers.js · React · Vite · Hono.js · Cloudflare Workers · Upstash Redis · Hardhat
Let's build something
extraordinary.
Get In Touch.
Send a message
Availability
Open to full-time roles
© 2026 Sur Vaghasiya