Rutgers '28 · B.S. Data Science & Economics · minor CS · Bangkok ⇄ New Brunswick

Chirath Soithong

I build automations for the repetitive, time-consuming work — so people get their time back.

At Agoda, summer 2026, I built two Slack tools for a commercial team. At home, a workflow reads me the news on weekday mornings. I interview the people first, and a person still makes the call.

2
tools built at Agoda, Jun–Aug 2026
95
experiments on Rutgers' GPU cluster
May '28
expected graduation

Looking for part-time or remote work in AI automation, data analysis, machine learning or consulting · NYC hybrid

Companies · worked at and with hover to pause · moves on its own
  • Agoda
  • Booking.com
  • Priceline
  • Booking Holdings
  • BuzzeBees
  • Omnisient
  • Rutgers
Tools I build with
  • Python
  • Slack
  • Telegram
  • n8n
  • Claude
  • Jira
  • GitHub
  • PyTorch
  • Databricks
  • Google Calendar
  • Google Sheets
  • Canvas
  • GitLab
  • Cloudflare
  • Inworld
  • Figma

Running now

two workflows · both running
  • Morning news briefing
    Each weekday morning, after the market opens: general, finance and tech news, summarised and read aloud, with links and pictures, sent to Telegram.
    running
  • Daily planner
    One Slack plan for my whole day, not just classes: coursework, calendar and anything I add by hand, re-sorted when the day changes and followed up until it is done. Reading my email is next.
    running

Green means it is running. I set this by hand and turn it grey if something breaks.

demo · the shape of the real Sep 16 message · text summarised · photos are illustrative, open-licence

Photos: Federal Reserve (public domain); U.S. Dept. of Energy (public domain) — illustrative

Selected work

grey = the old waycolour = what I built a person decides
01
Agoda · Bangkok ·
summer 2026
Live

Commercial checks, reviewed by a person

Does
Finds hotels that could also be sold through a partner site, checks the rules, and lists them in Slack with a ready-to-edit note.
Helps
A commercial manager who ran these checks by hand, and the account managers the note is drafted for.
beforeafter · built so no one hand-checks rows or starts a blank email
PythonSQLSlackhanded over
Read the story
02
Agoda · Bangkok ·
summer 2026
Live

Questions in one channel, answered or routed

Does
One Slack channel for colleagues' questions. A bot answers the common ones, looks things up, and opens a ticket when a person is needed.
Helps
A commercial team that got the same questions by email, Slack and Teams, each with a round of back-and-forth.
beforeafter · built so the same answer is not retyped
SlackJirateam interviews
Read the story
03
Personal · 2026 ·
built and run by me
Running · for me

The morning news, read to me

Does
Each weekday morning, after the market opens, it gathers general, finance and tech news, writes a short summary with links and pictures, and sends it as audio to my phone.
Helps
One listener: me. I wanted the news without scrolling several sites, and without making low-effort AI content.
beforeafter · listen instead of read, nothing to trigger
n8nClaudeInworld voiceTelegram
Read the story
04
Rutgers Quant Finance Club ·
spring 2026
Researchsubmitted to the Rutgers Quant Finance Club

Does a bigger model trade better?

Does
Rebuilt a published model that guesses the next price move from an order book, then tested simpler versions: 95 experiments on the university GPU cluster.
Helps
Anyone about to add complexity: higher accuracy did not reliably mean better simulated trading once costs were counted.
beforeafter · no more guessing whether complexity pays
PyTorchGPU clusterbacktests with costs
Read the story

How I work

the same five steps, every build
  1. 1
    Interview the people

    Who handles what, what a good answer looks like, where things get stuck.

  2. 2
    Map the workflow

    The steps, the rules, and the exact point where a person has to decide.

  3. 3
    Build with AI help

    AI for the typing; fixed rules wherever an answer must come out the same every time.

  4. 4
    Check it on real cases

    Run it, look at the output with the owner, log every wrong answer, fix, repeat.

  5. 5
    Hand it over

    Docs, a walkthrough, and enough understanding that the owner can change it without me.

Also built

one line each · ask me about any of them
  • Portfolio risk monitor — prices and ticker news in one pipeline to flag next-day downside risk; written up as a course report.
    course · 2025
  • Revenue prediction, BuzzeBees — Databricks SQL joining ad spend to sales data; prototyped a model for the return on boosting a partner's post.
    internship · 2025
  • Dip, campus ride-pooling — concept, Figma prototype and survey-backed pitch.
    pitch · 2025
  • Memory-game hackathon — simple browser games built with a team around memory-focused activities.
    hackathon · 2024