Korakrit Tonpongphan (Win)

Financecontract

Automated Alpha Research Pipeline

AI and genetic search propose trading signals; strict checks decide what is submitted

What it is

A Python pipeline, run with Claude Code, that automates alpha research on WorldQuant BRAIN, from first idea to submission.

What I built

• A sync job that keeps a local SQLite copy of the platform's catalog of operators and data fields, plus my past results.
• Two ways to propose candidates: a language model that builds ideas only from the verified catalog, fenced by a deterministic validator, and a genetic algorithm search.
• A checker that simulates each candidate on BRAIN and reads every pass or fail check from BRAIN itself, never from hardcoded limits, plus a filter that skips near-copies of past ideas.

Result

• Global Alpha Competition 2026: 1st in Singapore
• All Region Competition 2026: 1st in Singapore, 68th globally
• International Quant Championship 2026: Top 15 in Singapore, Stage 2

AIinternship

Building Manager Assistant

A chat assistant that answers, schedules and reports on live energy data for multi-property managers

Problem

Managers overseeing multiple buildings had to wait for each site's manager to send an energy report. I worked on the reporting pipeline so they could request reports directly and schedule recurring deliveries.

What I built

• Connected a Claude Agent SDK assistant to the company's live energy and sensor data through MCP tools.
• Chat-based scheduling: a manager describes a recurring report, and the agent runs it on fresh data and emails it.

ResearchIROP

Fraud Detection with Diverse Sampling

Diverse sampling beat random sampling, but neither beat using all the data

Question

An Isolation Forest trains each tree on a random sample of the data. Does choosing samples that are diverse on purpose, with a determinantal point process (DPP), catch fraud better?

What I did

• Benchmarked six setups on 10,000 synthetic card transactions (5% fraud) from the Synthetic Financial Fraud Dataset on Kaggle: full data, random rows, random features, DPP features, and two DPP–Isolation Forest designs.
• Built single- and two-stage DPP–Isolation Forest variants, the two-stage one adapted from Thakkar et al. (2024). Code on GitHub.
• Measured diversity directly: DPP subsets are more spread out (average pairwise distance 6.03 vs 5.91).

Result

At the same budget, DPP beat random sampling (fraud F1 0.19 vs 0.16), but no sampling method beat the full data (0.26). PCA showed why: fraud is explained almost entirely by one feature, the amount, which leaves little diversity to exploit.

Softwarecoursework

Hotel Booking Platform

A working hotel booking site, from search to payment. I built the backend.

The brief

Build a hotel booking web app for Ascenda's loyalty rewards programme to a set specification: destination search, hotel results with filters, room details, and booking with payment.

What I built

• Express API endpoints that connect the frontend to PostgreSQL and to Ascenda's hotel API, and shape each response for the frontend.
• Destination search from memory: exact matches in about 3 ms, with a fuzzy fallback for typos (about 90 ms) so "pariss" still finds Paris.
• The booking backend and database schema: each booking, its guest and its payment saved together in PostgreSQL, with the price checked on the server, never taken from the browser.
• Unit and integration tests for the search and booking layers.

Softwarehackathon

Bicycle Rack Network

2nd prize at a city hackathon for a live bike-rack map wired to real smart locks

Problem

Tampines bike racks are often over 75% full, and an estimated 25–50% of parked bikes are abandoned. Riders cannot find a free rack, and the town council cannot find the abandoned bikes to clear them.

What I built

• The rider app: a live map of about 325 racks from LTA open data, with routing and cycling navigation to the nearest free rack.
• Live lock status from the team's ESP32 smart racks over MQTT, so each lock and unlock shows in the app as it happens.
• The maintainer console: CCTV-detected illegal parking, abandoned bikes and full racks in one triage view. Code on GitHub.

Result

2nd prize at the Jom AI @ Tampines Hackathon.

Hardwarecoursework

Connect 3 on FPGA

A two-player game running as software on a Beta CPU our team built on an FPGA

The brief

SUTD's core computer-architecture course: study the Beta CPU architecture, build it on an FPGA starting from the course template, then run a game on it as software rather than as hardwired logic.

What I built

• The CPU's datapath and its control logic, a finite-state machine, in Lucid HDL.
• Testing on the real board: an automated ALU tester and unit tests per module, alongside the whole-CPU test from the course template, plus an on-board debugger that reads registers without recompiling. Code on GitHub.

Hardwareinternship

Foot Pressure Pipeline

A live foot-pressure heat map from a 16-sensor insole, with under 0.05% data loss

Problem

Commercial foot-pressure insoles cost from about US$3,000. The lab built a low-cost insole with its own circuit board, and needed a reliable way to stream its readings live for gait analysis.

What I built

• A new data pipeline: ESP32 firmware in C/C++ that sends all 16 pressure channels as JSON over MQTT.
• A Python receiver with a live view: a heat map of the foot and a time series of every channel.
• A reliability test that sends and counts 360,000 messages: 0.04% lost. Code on GitHub.

Result

Co-author of "Cost-effective integrated insole with mirrored PCB design for enhanced foot pressure monitoring", IS-PAIR 2025, published by Frontiers (abstract book, page 50).

ResearchISEF

Acoustic Streaming on a Triangular Vibrating Structure

2nd Grand Award at Regeneron ISEF 2024, among over 1,800 finalists from 70 countries

Question

A sharp-edged structure vibrating in a fluid drives a steady flow, called acoustic streaming. How does the angle of a triangular tip change that flow?

What I did

We simulated the flow with thermoviscous and Navier–Stokes equations, and measured it in experiments with particle image velocimetry (PIV), processed in MATLAB.

Result

• The jet changes direction, from outward to inward, as the apex angle widens, which prior research had not predicted.
• At large angles a sideways jet appears, and the angle also changes the number of vortices and the streaming speed.
• A body-force analysis, comparing the acoustic force with the viscous force, explains why. This helps design microfluidic devices that mix, circulate or trap tiny particles.