Scroll
Data Scientist · ML Engineer · Chicago, IL
I build ML systems that actually make it to production. Not "works on my machine" — real pipelines, real users, real results. I always track whether something actually worked.
Open to work
I'm a data scientist based in Chicago. I spent the last three years figuring out that building a good model is honestly the easier part — getting it into production, getting people to trust it, and proving it actually changed something is where most projects fall apart. That's the part I care about most.
I just graduated with my MS in Data Science from Illinois Tech. Before that I was a co-op at Labelmaster building demand forecasting systems, worked on churn prediction at August Infotech, B2B lead scoring at Orion Technolabs, and won a track at HackIllinois with Caterpillar. In every role I stayed involved from the first meeting to the final dashboard — I don't like handing things off mid-flight.
Outside of work I've tutored 50+ students in Python through Superprof, which turns out to be one of the best ways to actually understand something. I'm also a McKinsey Forward alumna and have three cloud certifications I actually use.
I'm currently looking for full-time data science or ML engineering roles. I'm on OPT STEM so I can start right away.
My Process
STEP 01
Frame
From vague asks to measurable bets.
I don't start modeling until I can write the decision this will change in one sentence. Business question first, dataset second. If the success metric isn't defined before I open a notebook, I define it — even if that means pushing back on the brief.
STEP 02
Prove
From model to rigorous evidence.
37-fold rolling-origin backtesting. 8-week A/B tests with proper holdouts. RAGAS evaluation at 0.92 faithfulness. I don't call it working until numbers say so — and when XGBoost outperformed LSTM across all departments, I documented it and changed the model. That's engineering, not ego.
STEP 03
Ship
From notebook to production trust.
MLflow tracking, FastAPI endpoints, CloudWatch monitoring, schema-validated outputs. The model is the last mile — the infrastructure is the real product. I don't call it done until it's measured in production and someone outside the data team is using it to make decisions.
Data Science Co-op
Labelmaster · Chicago, IL
Jan 2026 — May 2026Architected an end-to-end monthly sales forecasting system spanning 8+ departments. Transitioned from recursive to direct multi-output LSTM, eliminating error compounding from sequential predictions. Designed bias-correction layers and built a rigorous rolling-origin backtesting framework (37 folds) that revealed XGBoost outperforming LSTM across all departments — a finding that informed final model selection and saved weeks of tuning the wrong architecture.
Graduate Teaching Assistant
Illinois Institute of Technology · Chicago, IL · CS487 Software Engineering
Aug 2025 — May 2026Supporting 50+ students through weekly office hours, live sessions, and rubric design for research framework papers. Designed assessment criteria that improved student clarity on software design principles by standardizing evaluation across multiple project types.
Data Scientist Intern
August Infotech · Surat, India
Jan 2024 — Jul 2024Built a production churn prediction and expansion-likelihood platform for a B2B SaaS client on AWS. Engineered RFM-style features from 500K+ event logs, trained XGBoost achieving 0.81 AUC-ROC (vs 0.68 logistic baseline), and deployed batch + real-time scoring via Lambda. The key deliverable wasn't the model — it was the 8-week A/B test that proved model-driven onboarding actually moved revenue metrics at the company level.
Machine Learning Intern
Orion Technolabs · Ahmedabad, India
Jun 2022 — May 2023Built a B2B lead scoring system that improved top-decile conversion by 11–15% over the existing rule-based approach (AUC-ROC: 0.77–0.80). Compared logistic regression, random forest, XGBoost, and MLP before settling on XGBoost for its explainability tradeoff. Deployed via batch scoring on EC2 and a Flask REST API, integrating hot/warm/cold priority bands directly into the CRM so sales teams saw the output without touching the model.
Master of Science, Data Science
Illinois Institute of Technology
Aug 2024 — May 2026 · Chicago, IL
Focus: Applied ML · NLP · Agentic AI · Federated Learning · Statistical Experimentation
Google Developer Student Club
Bachelor of Engineering, Information Technology
Gujarat Technological University
2020 — 2024 · India
Foundation in computer science, algorithms, data structures, and software engineering principles.
Coding Club
ML Foundations Training
Amazon Web Services · Mar 2025
✓ Verified CredlyData Analysis Using Python
IBM · Jun 2023
✓ Verified CredlyData Visualization Using Python
IBM · Jun 2023
✓ Verified CredlyML with Python: Foundations
LinkedIn Learning · Mar 2023
✓ CertificateMcKinsey Forward Program
McKinsey & Company
✓ AlumnaGoogle Cloud Skills
Google Qwiklabs · Since 2021
✓ 13 Badges · View ProfileHover any card to see what's inside.
Flagship AI systems and ML projects built end-to-end
WanderMind AI
Multi-Agent RAG Platform · LangGraph · Neo4j · 3 Vector DBs
Adaptive RAG with fine-tuned Mistral 7B query router (94% accuracy, 60ms latency). Triple-layer constitutional validation. BERT personality classifier (91% accuracy) across 7 dimensions. Faithfulness: 0.76 → 0.92.
CatSense
Multimodal AI Inspection · HackIllinois 2026 · Caterpillar Track
Full-stack multimodal assistant with photo + voice input, RAG-grounded reasoning via Actian VectorAI, triple-layer hallucination control, schema-validated JSON (Zod) for deterministic UI. Deployed to Cloudflare Workers.
PulseBoard AI
A/B Testing & Causal Inference Platform · Bayesian + Frequentist
CUPED variance reduction (38% SE decrease), auto-diagnostics for 8 pathologies, sequential testing. Causal toolkit: DiD, Synthetic Controls, RDD, IV. Analysis time: ~2 hours → under 10 minutes.
TrustWeight
Asynchronous Federated Learning · PyTorch · ResNet-18
Momentum-based gradient projection filtering harmful stale gradients. Quality-aware aggregation with freshness functions. 52–53% accuracy on CIFAR-10 vs 11–28% for FedAsync/FedBuff baselines under 20–50% delayed clients.
Crypto Market Analyst Agent
LangGraph · GPT-4 · Autonomous Tool Selection
Autonomous LangGraph agent with GPT-4 reasoning and graph-based workflow. 3 custom tools for live price, volume, and sentiment data. Thread-based memory for cross-query context. Production safeguards and graceful fallbacks.
UK Real Estate Intelligence
25M+ Records · XGBoost · SHAP · K-means
End-to-end ML pipeline over 25M+ transactions. Ensemble models (RF, XGBoost) with SHAP analysis, K-means clustering producing 5 buyer segments (silhouette 0.68). RMSE £42K, R² 0.87. Spark-compatible for scale.
Open to full-time Data Scientist, ML Engineer, AI/GenAI Engineer roles starting 2026. Work-authorized on OPT STEM extension. If you're solving hard problems with AI — and you care whether the solution actually works in production — I want to hear about it.