Prince Kumar

Software + AI Engineer · College Station, TX

Computer Science at Texas A&M · Engineering Honors · Class of 2028

Email LinkedIn GitHub

Experience

Finch / ApplyEasy

2026 — Present

Full-Stack Developer

Building a job-search platform that brings role discovery, tailored application materials, and application tracking into one workflow.

  • SaaS
  • Ranking systems
  • Analytics
  • Cloud deployment

Maroon Fund

2026 — Present

Quantitative Developer

Working on time-series machine learning, backtesting, and risk-aware evaluation.

  • Time-series ML
  • Backtesting
  • Risk evaluation

Projects

Finch

A job-search startup that finds relevant roles, tailors application materials for each one, and keeps the whole search organized. We brought role discovery, ATS-ready résumés and cover letters, and application tracking into one workflow. Finch is live with ~200 weekly active users and thousands of internship and new-grad roles to browse.

  • Startup
  • Job search
  • Applied AI

grocerAlgo

I built grocerAlgo after spending too much time backtracking through H-E-B. It converts store-guide PDFs into walkable maps and turns a shopping list into an entrance-to-checkout route. A 25-item pilot cut the walk from 818m to 426m, with route updates in 216ms and 22 stores supported.

  • Route optimization
  • Computer vision
  • FastAPI

filterModel2016

A U-Net that learns the 2016 phone-selfie look from synthetic pairs and applies it to new photos. I built a parameterized 2016 filter in Python and synthesized perfectly aligned clean→filtered pairs on the fly, then trained a shallow RGB U-Net with L1 on that free supervision. Best validation L1 hit 0.021. The learned look — milky blacks, warm-magenta skin, soft glow — transfers to unseen portraits, buildings, and landscapes the network never trained on.

  • PyTorch
  • U-Net
  • Image-to-image

Arbitrage Engine

A C++ engine that compares Kalshi and Polymarket prices, spots mismatches, and sends both sides of a trade together. I built separate live feeds for Kalshi and Polymarket, then kept the comparison and trade decision path small enough to react immediately. The private engine watches both markets at once, checks the opportunity and risk, and sends paired orders when the prices line up.

  • Prediction markets
  • Low latency
  • Risk controls

SkinScan

A research pipeline that finds visible skin concerns, maps them to face regions, and recommends products through reviewed rules. I ran an A/B test on native-resolution tiles versus zoomed images, kept failed experiments in the record, and added visual checks before trusting aggregate scores. Tiling lifted recall from 4% to 70% in the serving test. The final detector reaches 0.743 recall, and product ranking uses the simpler baseline that beat the learned model.

  • PyTorch
  • Mask R-CNN
  • FastAPI

Synapse

A local-first coordination layer that gives coding agents shared context and warns them before incompatible edits waste the team’s time. Synapse lets an agent check active edits and changed contracts before it writes. Raw code stays local while the team shares only coordination signals, contract deltas, and resolutions. The working prototype includes a realtime daemon, deterministic conflict engine, GitHub reset flow, team briefings, and five tools agents can call directly.

  • TypeScript
  • WebSockets
  • SQLite

agentNotch

Live Claude Code, Codex, and Cursor sessions and usage limits, tucked into the MacBook notch. I used the physical notch as ambient status space and read existing local session data instead of adding another account, server, or Electron app. One lightweight macOS panel shows live sessions, limits, multiple accounts, and optional approval controls while keeping all activity on the machine.

  • Swift
  • SwiftUI
  • AppKit

Southwest Spoilage

A calibrated model that flags at-risk crew sequences before departure, when schedulers still have time to act. I audited leakage, split the data by time, used only pre-departure snapshots, and reported calibrated risk so schedulers could choose their own action threshold. The final blend reached 0.822 ranking AUC and won the TAMU ML Club competition. It also identifies severity reliably, not just whether something may go wrong.

  • scikit-learn
  • Time-series
  • Calibration

VectorDB

A handwritten C++20 vector database: SoA storage, exact top-k, checksummed snapshots, WAL with fsync and crash injection, then LSM memtable/segments/compaction — no Faiss, no wrappers. AoS vs SoA storage, id→position indexing, cosine/dot/L2, exact top-k heaps, SoA snapshots, WAL log-before-mutate with fsync and a fork crash matrix, then LSM flushes that write a batch instead of rewriting the whole photo. Simple correct version first. Exact search is the ground truth. Compaction merges segments instead of dropping old files. Metadata filters are next; HNSW waits until storage is one path.

  • C++20
  • WAL
  • LSM
  • SoA
  • Heaps

PiPartner

Photo-based math tutoring adopted by educators for classroom use, with 400+ downloads and a 5-star rating.

  • iOS
  • AI education

TAMUSkate

A campus community app with social maps, videos, live skate spots, and AI-assisted safety checks.

  • React Native
  • Community