Machine learning,
shipped like software.

A bitemporal financial data platform built around one question: how do you know the number is right? Point-in-time price storage, a dataset-level quality gate, and an AI analyst whose every numeric claim is checked against the tool output its prompt was built from. On its first run against a live scheduler, the quality gate caught a fabricated 9751% price move — two currencies silently sharing one database row — introduced by my own bug fix, root-caused and fixed the same day.

5 years of my own chess (4,635 games · 328,258 Stockfish-analyzed moves) treated as an ML problem. Five models compared with 5-fold CV, McNemar's test, ROC analysis — and a data-leakage fix that traded a flattering 78.21% for an honest 72.49%. The insights took my rating from 597 → 1,423.

Upload a table and a LangGraph agent cleans it, picks a target, trains and ranks models, and measures what drove the predictions — then refuses any conclusion the data can't support. Dual leakage detection caught Titanic's leaked alive column (MI 1.000) and abandoned a bike-share target that was just arithmetic (R² 1.000). LLM for judgement only; every metric is deterministic.

Ensemble (XGBoost + LogReg + RF) over 18 seasons of ball-by-ball data, 49 engineered features — Elo, phase stats, venue history. Walk-forward validation, 2,000-run Monte Carlo season simulation, P10–P90 player projections. Gave RCB the top title probability (25.4%) — and RCB won.

Terrain-routed kinematic-wave model reconstructing the 2026 Bhote Koshi–Trishuli flood from public satellite and DEM data. Validated at 3 independent gauges over 176 km (within ~10–16%). Monte Carlo uncertainty bands contain all observed arrivals. Human exposure estimated for ~38,200 people in the modelled corridor. Footprint underperformed a simple baseline — reported as a finding, not hidden.

Can semantic embeddings detect how an organization's public narrative evolved — without hand-defining the phases? Case study: OpenAI 2015–2026 (66 events). Cross-validated across GloVe and SBERT: both independently flagged the same 11-day window in November 2023 as the strongest detected boundary. 2/5 turning points in exact agreement across backends.

Can public environmental data predict road-accident risk in Indian cities? ~596K records from 453 CPCB stations + MoRTH data → PostgreSQL warehouse → 3×3 model ablation → per-prediction SHAP → FastAPI + Streamlit dashboard. Dockerized, CI, live-hosted. Its anomaly detector's first real catch was a faulty gauge reporting 20× actual rainfall — excluded and documented.

Published, peer-reviewed and awarded: real-time accident detection via sensor fusion + ML classification, neural network trained on 10,000+ samples, MQTT emergency-alert pipeline on Arduino. Best Paper at IC-CCDS 2025 (MULTICON-W), Mumbai.

Full-stack platform: adaptive psychometric assessments, Gemini-powered personalized roadmaps, automated PDF reports, Razorpay payments. RESTful API with 12+ endpoints at <200ms, scaled to 200+ concurrent users on Railway. 50+ profiles processed, 89% satisfaction.

Final-year project: a digital twin for industrial motors. Multi-layer fault-detection pipeline on synthetic MATLAB data, deployed as an ESP32 IoT system streaming over MQTT to a real-time dashboard with remaining-useful-life estimation. End-to-end latency under 500ms.

Can three independent public signals — Google Trends, Wikipedia pageviews, and YouTube creator activity — be fused into a measurable proxy for public attention, and does it actually beat chance? Permutation test (p=0.003) says yes; a holdout-by-event-type experiment (F1 collapsed to 0.000 on unseen categories) says where it fails. Both findings reported, neither hidden. 14-section Streamlit dashboard with signal lag, topic clustering, and bootstrap CIs.

Apple announced the iPhone Duo (its first foldable) — I wanted to know what consumers actually thought. 5,104 YouTube comments decomposed by Aspect-Based Sentiment Analysis across 12 product dimensions. Human-validated benchmark (175 hand-labelled comments, 5-fold CV) yielded F1=0.384 — the gap between that and the auto-label F1 of 0.959 is itself the finding. Sentence embeddings + HDBSCAN surfaced 19 topic clusters; bootstrap CIs on aspect sentiment with Chi-square comparison across both phones.

ESP8266 + HC-SR04 ultrasonic sensors monitor bin fill levels in real time, push to ThingSpeak cloud, and trigger GraphHopper route optimization for collection vehicles. ±1cm sensor accuracy at 30–60s update frequency; companion dual-app (driver route + public bin locator). Research project with 20–30% measured reduction in travel distance and 34% fuel savings over traditional fixed-schedule collection. Selected for the NCC Idea & Innovation Camp — Republic Day 2026.