Research Engineer @Provue • ML • Data Science • Building systems, not just models

I build machine learning systems that solve real problems. My approach: deep understanding of data → rigorous validation → systems people can actually trust. I believe honest metrics matter more than impressive benchmarks.

🌐 Portfolio • 📄 Resume • 💼 LinkedIn • 📧 Email

• 🏆 Best Paper Award — MULTICON-W 2025
• 🎓 Dean's Scholar — CGPA 9.5/10
• 💼 Research Engineer — Provue (Sep 2026 → Present)
• 🤖 13 End-to-End AI/ML Systems — From conception to production
• 📊 596K+ Environmental Records Analyzed — Across 453 CPCB stations; identified a faulty rainfall sensor
• 🔬 Research Intern — Wadhwani Foundation (Sep–Dec 2025)

• RL environments and evaluation systems for frontier AI agents at Provue
• Production ML systems with FastAPI, PostgreSQL & Docker
• LLM agents with grounding and claim verification, not just chat wrappers
• Explainable AI using SHAP and model interpretation

1. MarketIQ — Bitemporal financial platform with a self-verifying AI analyst

• On its first live run, its own quality gate caught a bug I'd just shipped: a fabricated 9751% price move traced to two securities silently sharing one database row
• AI claim verification tags every numeric answer as supported / derived / contradicted / unsupported against real tool evidence
• 460 automated tests, CI-gated on every push

2. ChessIQ — Chess outcome prediction, and the leaks that faked it

• Found and fixed 3 data leaks: post-game engine metrics, post-game Elo written into "pre-game" features, and a mislabeled target
• After removing the leaks: AUC 0.5001 — essentially chance, exposing the previously inflated 78.21% result
• The leak-hunting and validation process became the actual deliverable — not the model

3. Autonomous Data Analyst — Agentic CSV analyst built with LangGraph

• Cleans, profiles, models, and explains CSV datasets — and explicitly says when the data cannot support the question
• Caught 5 classes of silent leakage/failure across 9 public datasets, including Titanic's alive/survived duplication
• 214 automated tests, with zero network calls required to pass them

4. IPL 2026 Prediction — Cricket season prediction & simulation

• 1,175 matches / 279K deliveries from 2008–2026 with 49 engineered features
• XGBoost + Logistic Regression + Random Forest ensemble with 52–55% walk-forward accuracy
• Reported the result as the realistic prediction ceiling for T20 outcomes from public data
• 2,000-run Monte Carlo simulation → RCB favored at 25.4% to win IPL 2026

5. CASCADE — Terrain-routed flood simulation & human exposure model

• Reconstructed the 2026 Nepal Bhote Koshi–Trishuli flood from satellite and DEM data; validated arrival times at 3 independent gauges across 176 km to within ~10–16%
• Monte Carlo analysis placed all 3 observed times inside the P5–P95 uncertainty band
• Honest negative reported: flood footprint model (IoU 0.18) did not beat the geometric baseline (0.27) — documented, not papered over

6. TRACE — Temporal narrative analysis via semantic embeddings

• Unsupervised detection of narrative phase shifts across OpenAI's 66 public events (2015–2026) without pre-defining phases
• Cross-validated across GloVe and SBERT independently — 2/5 turning points exact match, third pair within 11 days
• No finding claimed unless it survived both backends

7. CivicLens — Road accident risk intelligence for Indian cities

• 596K+ environmental records across 453 CPCB stations, organized in a layered PostgreSQL warehouse
• 3×3 ablation study showed environmental signal is real but secondary to accident history
• Anomaly detection identified a faulty rainfall sensor (TN004) during the build

8. GTA VI Hype Intelligence — Multi-source public attention detection

• Fused Google Trends, Wikipedia pageviews, and YouTube into a Hype Index; permutation test p=0.003 vs random baseline
• Three generalization experiments run — holdout by event type collapsed to F1=0.000 on unseen categories; reported equally
• Trends + Wikipedia alone (F1=0.599) beat all three signals combined

9. Apple Effect — Product perception intelligence

• 5,104 YouTube comments decomposed via ABSA across 12 product aspects during iPhone Duo launch week
• Human-validated benchmark F1=0.384 vs auto-label F1=0.959 — the gap is the finding
• 19 topic clusters via UMAP + HDBSCAN, no supervision

10. Smart Helmet IoT — Real-time accident detection system

• 🏆 Best Paper Award — MULTICON-W 2025
• 94% fall-detection accuracy with Pushbullet + GSM emergency alerts and live GPS

Languages: Python, SQL, C++, JavaScript/TypeScript

Machine Learning: XGBoost, scikit-learn, PyTorch, SHAP, Ensemble Learning

Agentic AI: LangGraph, Gemini, RAG, Claim & Grounding Verification

Backend & APIs: FastAPI, Flask, PostgreSQL, SQLAlchemy, Docker

MLOps & Cloud: Apache Airflow, GitHub Actions, Google Cloud Run, Railway, AWS SageMaker

Data & Visualization: Pandas, NumPy, Plotly, Streamlit

Hardware & IoT: ESP32, ESP8266, Arduino, MPU6050, NEO-6M GPS

Sep 2026 – Present

• Building and iterating on realistic, resettable environments for frontier AI agents to interact with tools and real-world scenarios
• Designing automated verifiers that programmatically determine whether an agent has completed a task correctly
• Investigating failure modes in agent behaviour; contributing to task design, reward mechanisms, and evaluation methodology to make experiments reproducible and measurable at scale

Feb–July 2026

• Engineered a semantic retrieval pipeline for RAG-based LLM applications
• Reduced manual assessment time by 70% while maintaining 95%+ prediction consistency
• Deployed FastAPI inference on Railway with sub-500ms latency across 500+ profiles

Sep–Dec 2025

• Analyzed 10,000+ entrepreneurial records using Pandas and NumPy
• Surfaced 12+ market trends through structured data analysis
• Improved data quality from 62% → 88%
• Authored 15+ quarterly research reports

2024–2025

• Managed 40+ contributors and coordinated publication of 100+ page quarterly issues

2022–2026

• Represented and coordinated 60+ engineering students across 4 years

MULTICON-W 2025 — 16th International & National Conferences and Workshops

🏆 Best Paper Award

• MPU6050-based fall detection with GPS location tracking
• Pushbullet and GSM-based emergency alerting
• Real-time accident detection and notification architecture

B.Tech in Internet of Things — Thakur College of Engineering & Technology

• Graduated: June 2026
• CGPA: 9.5/10 · Dean's Scholar
• Best Paper Award — Smart Helmet for Accident Detection, MULTICON-W 2025

• Practical Data Science with Amazon SageMaker — AWS
• Data Analytics Job Simulation — Deloitte Australia (Forage)
• Machine Learning Onramp — MathWorks

• ✅ Honest > Impressive — Report real metrics, even when they hurt
• ✅ Data > Fancy Models — Better data beats unnecessary complexity
• ✅ Understanding > Optimization — Know why the system behaves the way it does
• ✅ Systems > Notebooks — Build for deployment, validation, and monitoring
• ✅ Validation is non-negotiable — Find leakage, test assumptions, report reality

📧 keyurchauhan15@gmail.com — always happy to talk about ML systems, agent evals, or data science.

"The difference between a model and a system is discipline: validation strategy, honest reporting, and systems people can actually trust."