Prince Kumar

Software & AI engineer. Computer Science at Texas A&M.

Incoming at Deloitte.

Experience

Deloitte

Jun 2027 — Aug 2027

Incoming Software Engineer · Austin, TX

Formeon AI

Feb 2026 — Present

Founding Engineer · San Francisco, CA

Building an AI decision-memory platform.

More about this role
  • Raised top-5 retrieval hit rate from 62% to 87% on 50 labeled decision queries by combining vector search, keyword search, and recency weighting.
  • Built ingestion from Gmail, Slack, Jira, Confluence, and Drive: 120K+ embedded chunks across a 14-table PostgreSQL schema, with p95 search under 150 ms.
  • Shipped eight Edge Functions for OAuth, webhooks, and incremental sync, cutting re-embedding cost by 85%.

TypeScript · PostgreSQL · pgvector · Supabase · RAG

Chaser

Oct 2026 — Present

Backend Engineer · College Station, TX

Building the backend for live bar queues and payments.

More about this role
  • Replaced multi-step queue writes with seven transactional PostgreSQL functions. Row locks prevent expiry jobs from interrupting an offer being accepted.
  • Integrated Stripe Connect deposits, platform fees, and idempotent PaymentIntents to prevent double charges.
  • Covered routes, state transitions, payment failures, and error fallbacks with 247 passing Jest and Deno tests.

TypeScript · PostgreSQL · Supabase · Stripe · Jest

Finch

May — Aug 2026

Software Engineering Intern · Remote

Shipped services for a job-search SaaS with 2K+ weekly active users.

More about this role
  • Raised top-3 job-match relevance from 71% to 93% on a 60-student evaluation, while reducing match latency from about 120 s to 50 s.
  • Built a distributed crawler for 10 ATS platforms with four parallel workers, rate limits, and transactional writes. Its first run discovered 12K+ career boards.

Python · Flask · React · AWS EC2 · PostgreSQL · CI/CD

Projects

Trading bot architecture: Kalshi and Polymarket feeds pass through opportunity detection and risk checks to order execution
  • C++20
  • WebSockets
  • Boost
  • OpenSSL
  • simdjson

Trading bot

A C++20 arbitrage engine for Kalshi and Polymarket. It ingests both markets’ feeds, compares complementary YES/NO prices after fees, and manages order execution and hedging.

Separate feed handlers, fee-aware opportunity detection, position tracking, and exposure-based quote limits keep market data, execution, and risk controls distinct.

Synapse architecture: coding agents connect through local daemons to shared coordination state
  • TypeScript
  • Python
  • Redis
  • PostgreSQL
  • MCP

Synapse

Realtime coordination for teams of coding agents. Synapse checks for conflicting edits before an agent writes, and shares contract changes across machines without sending raw code.

Published on npm, with stateless servers, Redis pub/sub, PostgreSQL, and peer-lock leases for coordinating concurrent edits.

SkinScan visual evaluation comparing zoomed crops with native-resolution tile detections
  • Python
  • PyTorch
  • Mask R-CNN
  • MediaPipe
  • FastAPI

SkinScan

A computer-vision research pipeline for identifying skin lesions and mapping them to face regions. A 10-class SA-RPN detector reaches 0.743 recall and 0.499 mAP@0.50 on held-out AcneSCU crops.

Native-resolution tiling improved serving-test recall from 4% to 70% versus zoomed crops. The project also explores product recommendations, with eligibility checks and explicit deferral when reviewed policy is unavailable.

grocerAlgo app showing a shopping list and its numbered route on an H-E-B floor plan
  • Python
  • FastAPI
  • OpenCV
  • SciPy
  • Docker

grocerAlgo

Turns a grocery list into an optimized walk through the actual store. It matches live H-E-B products to floor plans, groups aisle stops, and draws the route from entrance to checkout.

A 25-item pilot cut walking distance from 818 m to 426 m, a 48% reduction, with a 216 ms solve. The map pipeline converts store-guide PDFs into walkable graphs and validates product placement before offering a store.

VectorDB architecture showing its public API, memory layout, snapshots, write-ahead log, and LSM segments
  • C++20
  • LSM storage
  • WAL
  • CMake
  • CTest

VectorDB

A handwritten C++20 vector database built to understand storage and search from the ground up. Includes exact top-k search, cosine/dot/L2 similarity, metadata filters, and an LSM storage engine.

Built immutable segments, memtables, compaction, and a manifest, alongside a legacy snapshot-and-WAL path with fsync and crash-injection tests. The repository includes benchmarks and step-by-step engineering notes.

Other projects (3)

agentNotch

A native macOS panel for coding-agent sessions and usage limits, tucked into the MacBook notch.

Swift · SwiftUI · AppKit

filterModel2016

A U-Net trained on synthetic image pairs to recreate the 2016 phone-selfie look. Best validation L1: 0.021.

PyTorch · Image-to-image

Southwest Spoilage

A calibrated crew-sequence risk model with 0.822 ranking AUC. Winner of the TAMU ML Club competition.

scikit-learn · Time series

Open source

headroom

PR #647

Migrated Copilot reasoning models to the Responses API in an LLM-compression proxy.

CloakBrowser

PR #352

Fixed leaked Chrome processes by reaping them after the last CDP client disconnects.

More on GitHub

Skills

Languages
C++, Python, C, TypeScript, JavaScript, SQL
Backend & systems
PostgreSQL, Redis, Supabase, AWS, Docker, WebSockets, FastAPI, Flask, CI/CD
AI & ML
PyTorch, OpenCV, scikit-learn, XGBoost, pgvector, RAG, LLM APIs, MCP
Frontend & testing
React, Jest, CMake/CTest, PostHog, A/B testing

Education

Texas A&M University

May 2028

B.S. Computer Science · Engineering Honors · 3.83 / 4.0 GPA

Writing

All notes