Hi, I'm Arnav.

I build AI agents people use every day.

I'm a final-year CS student at NUS who builds AI products for people who aren't engineers, from safety officers at an energy company to stroke doctors and M&A analysts. Highlights include an agent officers use daily and a browser agent with 1,000 users.

See the work→ Ask the agent about me→

Right now I'm training a 1.38B-parameter language model from scratch for my final-year project. I'm open to full-time roles from June 2027, and part-time internships now.

Arnav Goel
me

What I've built

What it does

Officers ask a question and get a sourced answer in under 3 seconds. The agent reads the database, runs queries and searches records until it finds the answer, and learns from its past mistakes.

My part

I built the agent, the chat interface, the AWS deployment and the nightly tests, working directly with the officers who use it. I also proposed the voice version after seeing that officers on site couldn't stop to type.

Proof

A nightly run checks it against 20 questions with known answers, and 121 automated tests back it up. Questions that took an analyst an afternoon now take seconds.

Full case study →

What it does

You highlight anything on a page to ask about it, or give it a task and it clicks, types and navigates the site for you. Your keys and pages stay in your browser.

My part

I built and run it solo, from the first line of code to the store listing, including the check that confirms each action worked before the next one.

Proof

I raised the accuracy of telling a question from a request to act from 68.0% to 94.7%. It asks for confirmation before any purchase or payment. 149 tests.

Full case study → Chrome Web Store ↗ GitHub ↗

What it does

The platforms replace the team's manual compliance, delivery and procurement work across 6 university systems with one view that records every change.

My part

I watched how the team worked before designing anything, then built the login and access system and the engine that syncs with all 6 systems. We rolled each platform out in phases so the live network kept running.

Proof

They're in production with university single sign-on, a full audit trail, a stop button for faulty bulk changes, and 167 backend tests before every release.

Full case study →

What it does

Patients find targets in 3D by looking and pinching, while a clinician runs the session from an iPad and gets a PDF report at the end.

My part

I was the only developer on both apps, and rebuilt them over seven months of clinic visits and doctor feedback with NUS's 3D Design Lab.

Proof

Tested with 30+ stroke patients across two hospitals. I co-authored a paper submitted to ISMAR 2026.

Full case study →

What it does

8 agents each cover one area (financials, legal record, leadership, news and risk) and hand their findings to one shared report.

My part

I designed the planner, the parallel agents and the check that decides what goes into the report.

Proof

It drops any finding it can't trace to a source. Open source, 45 tests in CI.

Full case study → GitHub ↗

What it does

A second model re-checks only the answers where the first one's confidence is split, 55 of 10,570, and leaves the rest alone.

My part

I did the fine-tuning, designed the evaluation and built the confidence check on my own.

Proof

That raised the SQuAD score from 84.28 to 84.40 EM without touching answers that were already right.

Full case study → GitHub ↗

Where I've worked

Jul - Aug 2026

AI Engineer Intern

iEnergy Digital · built alongside the safety officers who use it

Built the safety agent, its nightly tests and the voice version, working with the officers who use it. Case study

Jun - Aug 2026

Undergraduate Teaching Assistant

National University of Singapore

Taught 20+ students in NUS's intro AI and machine learning course, and ran revision for 3 other courses.

Jan - Jun 2026

Software Engineer Intern

NUS Information Technology · Network Automation

Shipped 3 platforms for the university network team, including the login system and the sync engine. Case study

Apr 2025 - Jan 2026 · alongside NUS IT

Software Engineer & Undergraduate Researcher

Interactive 3D Design Lab, NUS · with NUH and Apple

Built the Vision Pro stroke test and its iPad control app as the only developer, over seven months of clinic visits. Case study

Sep - Dec 2025

AI Developer Intern

Klimacap · built alongside the M&A analysts who write the reports

Cut M&A analysts' report writing from hours per company to minutes with an AI pipeline, as the only engineer. Role details

May - Jul 2025

Software Engineer Intern

Aurionpro Solutions · SmartLender, a lending platform inside commercial banks

Shipped lending features that commercial banks use on SmartLender, and standardised releases across 3 environments so banks tested the exact build they'd run. Role details

Apr - May 2025

Lead Software Developer

Source Academy, NUS

Ran the CI/CD for the platform NUS computing courses use, fixing broken pipelines and reviewing outside contributions. Role details

Aug - Dec 2024

Research Assistant

National University of Singapore · data collection and analysis

Collected 200,000+ football-player records with web scrapers and cleaned them for modelling, then did the same for Twitch stats.

Side projects

What it does

I'm extending nanochat through pretraining, fine-tuning and RL alignment, supervised by Assoc Prof Eric Han. My baseline model scores 0.2603, the GPT-2-grade target.

How

The reference setup assumes 8 GPUs. I had 2, on a cluster that stops every job after 3 hours, so I made the jobs resubmit themselves and still reached 79% of ideal speed per GPU. Two runs on different hardware agree within 0.54%.

What it does

You describe the app you want. It turns that into a spec, splits it into files, and has several AI coders write them in parallel without stepping on each other.

How

Nothing ships until the app installs and builds, and a QA pass fixes what breaks. Follow-up requests only change the files they need to. It runs on free-tier Gemini or offline in demo mode.

repo ↗

What it does

OpenAI's Whisper gets a lot of Singaporean speech wrong. I fine-tuned its smallest version, whisper-tiny, on local speech, training just 1.4% of the model.

How

Errors fell from 66.2% to 23.8%, beating the next model size up at a fraction of the cost. The speech data is licensed, so the repo has the code and results but not the audio.

repo ↗

What it does

You give it a Singaporean health-insurance policy and a claim, and it works out the payout.

How

I trained just 0.032% of the model's weights, raising exact-payout accuracy from 46.7% to 73.3% and halving the average error, from $1,892 to $823.

repo pending publication

What it does

One interactive message controls everything: queue, effects, volume, live lyrics, and Spotify playlists pulled in automatically.

How

The DJ learns from what people skip and what they let play, and mixes similar tracks with some variety. 39 tests in CI, MIT licensed.

repo ↗eval demo ↗

What it does

Each agent searches live, then they agree on one itinerary that fits the budget, using real flight data.

How

I built the React front end and a Django API with login, with Gemini doing the planning. CI runs the tests and a production build on every push.

repo ↗

What it does

It pulls TikTok, Shazam, YouTube, Apple Music and curated playlists into one ranking, and shows why each song scored as it did.

How

The weekly replay only uses data available that week, so it can't use future charts to look good. Live demo on GitHub Pages.

repo ↗live demo ↗

What it does

It prices 5-minute prediction markets with option-pricing models on a few years of market data, and replays any time window exactly.

How

Every number comes with error bars and corrections for testing many ideas at once. 174 tests run in CI, and there's a read-only dashboard on GitHub Pages.

repo ↗live demo ↗

What it does

I contributed to the rebuild of the official NUS Students' Union site.

How

The whole student body uses it.

live ↗

The rest of my week

Leading

Cluster Leader & Fire Warden, PGPR

2025 - present

I lead 15 residents on my floor, running events, settling disputes and handling fire drills.

Teaching

Volunteer Tutor, Teach SG (MOE)

2026

Organising

Organiser & Mentor, Spatial Hack AI 2026 with Apple

7 to 10 Aug 2026 · Apple Developer Center Singapore

I organised and mentored a Vision Pro hackathon with Apple for 126 participants from NUS, NTU, SUTD, SMU and Imperial.

Say hi.

I graduate in June 2027 and I'm looking for full-time software and AI roles.