Hi, I'm Arnav.

I build AI agents that people actually use.

Final-year computer science at NUS. Over the last 18 months I've shipped for safety officers, stroke clinicians, M&A analysts and NUS's own IT team. One question drives all of it: does this work for the person using it, and can I show that it does.

See the work→ Ask the agent about me→

Right now: final-year research on LLM training at NUS. Open to full-time roles from June 2027.

Arnav Goel
me

Things I've built that people use

What it does

Safety officers at an energy operator ask a question in plain language and get a sourced answer in under 3 seconds. Behind it, an agent that reads the database, runs queries and searches records over up to 8 steps, and remembers its own mistakes so it gets better with use.

My part

Built the whole thing as the AI engineer intern, sitting with the officers who use it: the agent, the nightly evaluation, the deployment on AWS, and then a voice version that talks back in 30+ languages.

Proof

Graded every night in production on 20 golden questions against live ground truth, with 121 automated tests behind it. A question that used to cost an analyst an afternoon now takes seconds.

Full case study →

What it does

Highlight anything on a page and ask about it, or hand the browser a task and it clicks, types and navigates until the job is done. Local-first: your keys and pages never leave the browser.

My part

Built and maintained it solo from the first line to the store listing, including the agent loop that checks the page really changed after every action.

Proof

1,000 users on the Chrome Web Store and a Featured badge. The address-bar router tells an answer from an action correctly 94.7% of the time, up from 68.0%. 149 tests.

Full case study → Chrome Web Store ↗ GitHub ↗

What it does

Three internal platforms the university's network team runs on, replacing manual compliance, delivery and procurement work across 6 enterprise systems with one audited view.

My part

Software engineer intern across all three: the login and access system, and the sync engine that talks to the six external systems. Scoped it by watching the team work, then rolled it out around a live network.

Proof

In production for the network team. One login for everything through university SSO, long jobs engineers can cancel mid-run, and 167 backend tests run before every release.

Full case study →

What it does

A clinical test for spatial neglect after stroke, rebuilt for Apple Vision Pro: patients find targets in 3D space by looking and pinching, while a clinician controls the session from an iPad.

My part

Sole developer of both apps over seven months of clinic visits with NUS's 3D Design Lab, NUH clinicians and Apple; also in the room for testing, analysis and the paper.

Proof

Piloted with 30+ stroke patients in two hospitals. Captures how a patient searches space, which the paper test cannot. Co-authored an ISMAR 2026 submission.

Full case study →

What it does

Give it a company name and 8 agents research it in parallel, across financials, legal record, leadership, news and risk, then assemble one report where every claim links to its source.

My part

Designed and built the orchestration: the planner, the parallel agents, the provenance tracking, and the gate that decides what is allowed into the report.

Proof

Anything it cannot trace to a source is dropped, so a gap shows up as a gap instead of a confident guess. Public, MIT licensed, 45 tests in CI.

Full case study → GitHub ↗

What it does

A question-answering model that knows when it is unsure. 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

Individual project: the fine-tuning, the evaluation design, and the confidence gate.

Proof

Of the 55 answers it touched, it fixed 14 and broke 1, lifting the SQuAD score from 84.28 to 84.40 EM. Small on purpose: restraint is the result.

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 and shipped the safety and compliance agent, the nightly evaluation and the voice layer, sitting 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 module: search, logic, supervised learning, neural networks. Runs revision sessions for three other computing courses.

Jan - Jun 2026

Software Engineer Intern

NUS Information Technology · Network Automation

Shipped three platforms into production for the team running the university network, including the authentication stack 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

Sole developer of the Vision Pro stroke assessment and its clinician control app, through 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 investor-report assembly from hours per company to minutes as the only engineer in the firm: a staged pipeline researches the company and delivers a formatted report with every claim cited. Role details

May - Jul 2025

Software Engineer Intern

Aurionpro Solutions · SmartLender, a lending platform inside commercial banks

Parameterised Oracle PL/SQL deployments so a release behaved identically across 3 environments, development, SIT and UAT; shipped Apache Spark jobs and product features. Role details

Apr - May 2025

Lead Software Developer

Source Academy, NUS

Owned production CI/CD for the platform NUS computing courses run on: triaged failing pipelines to root cause, reviewed and merged external contributions, mentored contributors through releases. Role details

Aug - Dec 2024

Research Assistant

National University of Singapore · data collection and analysis

Built Selenium and Beautiful Soup pipelines that collected 200,000+ football-player records, then validated and reshaped them into model-ready datasets; did the same for Twitch channel and streamer statistics with Pandas.

Side projects

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.

Why it's interesting

Nothing counts as done until the app actually installs and builds; a QA pass fixes what breaks. Follow-up requests change only the files they need to. Runs on free-tier Gemini or fully offline in demo mode.

repo ↗

What it does

OpenAI's Whisper gets most Singaporean speech wrong. I fine-tuned the small version on local speech, training only 1.4% of the model.

Why it's interesting

Errors fell from 66.2% to 23.8%, better than the next model size up at a fraction of the cost. The speech data is licensed, so the repo ships the code and results, never the audio.

repo ↗

What it does

Give it a Singaporean health-insurance policy and a claim, and it calculates the payout, the kind of reading most people give up on.

Why it's interesting

Training only 0.032% of the model's weights took exact-payout accuracy from 46.7% to 73.3% and halved 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.

Why it's interesting

The DJ learns from what people skip and what they let play, and picks the next track with a mix of similarity and variety. 39 tests in CI, MIT licensed.

repo ↗eval demo ↗

What it does

Flight, hotel and attraction agents each search live and negotiate one itinerary that fits the budget, with real flight data behind it.

Why it's interesting

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

repo ↗

What it does

Pulls TikTok, Shazam, YouTube, Apple Music and curated playlists into one ranking, and shows exactly why each song scored what it did.

Why it's interesting

Recommendations stay fresh without being random, and the weekly replay only ever uses what was knowable that week, so the results cannot cheat. Live demo on GitHub Pages.

repo ↗live demo ↗

What it does

Takes a few years of market data, prices five-minute prediction markets with option-style models, and replays any window exactly so a result can be reproduced later.

Why it's interesting

It is honest about uncertainty: every number comes with confidence intervals and multiple-test corrections. 174 tests run in CI, and there is a read-only dashboard on GitHub Pages.

repo ↗live demo ↗

What it does

Contributor to the official NUS Students' Union site rebuild.

Why it's interesting

It is live at nussu.net and used by the whole student body.

live ↗

The rest of my week

Leading

Cluster Leader & Fire Warden, PGPR

2025 - present

I look after 15 residents on my floor at NUS: welcoming new arrivals, running floor events, sorting out disputes, and being the one who answers when the alarm goes off.

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

Helped run the Vision Pro hackathon and mentored the teams, 126 participants across NUS, NTU, SUTD, SMU and Imperial.

Say hi.

Graduating June 2027, open to full-time software and AI roles from then.