Skip to content
Ruhan Pandit
← Back

Projects

futures-data-store

The market data layer for Edge Lab, a futures research platform I'm building one package at a time, starting with clean, queryable historical data.

  • Ingests six years of 1-minute ES and NQ bars from Databento (about 4.2M rows, 2019 to 2024), normalizes them to Parquet partitioned by symbol and year, and queries them through DuckDB
  • Built-in data-quality checks for duplicates, gaps, half-day sessions, and contract rolls, covered by 30 tests (21 on synthetic data, 9 against the real dataset)
  • Returns bars as C-contiguous NumPy arrays so the C++ backtest engine coming next can consume them directly

Python, Databento, Parquet, DuckDB, pandas, NumPy, pytest

Ledge

A trade journal for futures that I built because the existing ones weren't customizable enough.

  • Full-stack app with a dashboard, trade log, strategy definitions, and a custom tag system for setups and mistakes
  • Charts of performance by strategy and tag to spot patterns across trades

React, TypeScript, Vite, Tailwind, Recharts, Node.js, Express, SQLite

YAC Over Expected

Built at the NFL Big Data Bowl NYU Edition (AWS x NFL) on a 3-person team, where I was the sole developer. A metric that separates how many yards after the catch a receiver creates from how much open field the defense gave him.

  • Joined NFL Next Gen Stats tracking data (122 games, 2021 weeks 1 to 8) with nflverse play-by-play to model yards after catch on 4,400+ completions
  • Simulates a race between the receiver and every defender at the moment of the throw to measure “YAC Available,” the open field before a defender can close
  • On short throws, explains about 40% more of YAC than nflverse’s expected-YAC model (R² 0.093 to 0.129)
  • Ranks 117 WRs, TEs, and RBs by YAC over expected; Ja’Marr Chase, Chris Godwin, Jonathan Taylor, and Christian McCaffrey led

Python, pandas, NumPy, scikit-learn, matplotlib

Agent-Writable Budget Tracker

A private budget tracker that two people and two AI agents use every day. The agents log card transactions through a role-based REST API, and both users see shared and per-person spending against monthly budgets.

  • Server-enforced permissions: each agent’s API key is scoped to one person, with the person derived from the key and never from the request body
  • Credential design: PBKDF2 for passwords, SHA-256 for high-entropy API keys, hashed session tokens, timing-safe comparisons
  • CSRF protection (Origin + Content-Type checks) and DB-backed login rate limiting, since in-memory limits don’t work on serverless
  • Idempotent writes via external_id, integer-cent accounting, and SQL migrations
  • Chose Turso over Netlify Database after modeling credit usage, to avoid pausing the main portfolio site

Astro, Netlify Functions, Turso (libSQL), TypeScript

Humana Resolve

Built with a team of 5 at the Humana Hackathon: a multi-agent assistant with a voice interface that answers healthcare member questions about claims, benefits, and coverage.

  • Root orchestrator routes each request to specialized agents for claims, benefits, privacy (Release of Information), and compliance, then combines their results into one plain-language answer
  • Voice interface: browser audio is transcribed with Google Cloud Speech-to-Text, and replies are read back with edge-tts

Python, Google Agent Development Kit, Gemini 2.5 Flash, FastAPI, Google Cloud Speech-to-Text, edge-tts

NYU Metaverse for Education

A fully interactive VR replica of the NYU Tandon MakerSpace, built with a team of 20+ students through NYU's Vertically Integrated Projects program.

  • Wrote custom Unity editor tooling in C# to assign PBR materials across 180+ meshes instead of setting each one by hand
  • Built a VR digital twin of the Tormach PCNC 440 mill: grip the machine and a state-machine sequence opens the doors, plunges the spindle, and moves the table

Unity, C#, URP, XR Interaction Toolkit

Folia

A portfolio analysis tool that reads brokerage screenshots and scores the portfolio on metrics like diversification.

  • Runs OCR entirely in the browser, so no data leaves the user’s device and there’s no backend to host
  • Deployed on GitHub Pages

JavaScript, Tesseract.js  Live

Housing Market Predictor

A machine learning model that forecasts U.S. housing market trends.

  • Trained on the House Price Index, S&P data, and U.S. Department of Housing and Urban Development data

Python, TensorFlow, NumPy, Pandas