Singapore · NUS Computer Science · Grad. May 2027

I build software where AI, operational systems, and human workflows meet.

Creator of Select to AI, a Chrome extension used by roughly 1,000 people✓ verified · public store listing. Shipped network-automation platforms inside NUS IT under real security constraints. Built an Apple Vision Pro clinical assessment with NUS and NUH, piloted with 30+ stroke survivors~ preliminary · ISMAR 2026 submitted.

View selected work

currently — AI Engineer Intern at iEnergy Digital · teaching NUS's intro AI/ML module

~1,000

Chrome Web Store users on Select to AI, with Featured status✓ verified · public listing

167

Backend tests validating 6 integrated enterprise systems at NUS IT✓ verified · source

30+

Stroke survivors in a Vision Pro clinical pilot with NUH~ preliminary · submission under review

The AI part, made honest

Don't scroll. Ask.

This console runs the same plan → retrieve → verify → answer pattern I built for Select to AI's browser agent and GUS's validation gate — on my own site, grounded only in the resume and case studies below. If it can't cite a source, it says so.

agent://ask-arnav · deterministic · fail-closed · nothing leaves this page

// pick a question, type your own, or start the guided tour below

>

// or tell the agent who you are, and it will drive the page for you:

The quirky one

Paste your job description. The page rewrites itself.

Drop in a real JD: the page extracts requirements, scores my evidence against each one, reorders the case studies to lead with what you care about — and tells you plainly what I don't have. Runs entirely in your browser.

// awaiting input — no JD analysed yet

Selected work

Work worth your three minutes

Four flagships, and one for the researchers. Each expands here — or opens as a full case study answering the same eight questions.

Problem

Jumping between a page and an AI chat tab breaks flow — copy, switch, paste, switch back. People wanted the AI brought to the selection, not the other way around.

My role

Sole builder and maintainer, end to end — permissions design, content-script injection, provider integrations, and the agent layer.

Constraints

Manifest V3's restricted permission model; users' API keys and page contents stay local, never routed through a server I control.

What shipped

Grew to roughly 1,000 users✓ verified · public listing and Chrome Web Store Featured status. The v3 agent plans and validates cross-tab tasks under exact-origin permissions, local API keys, and explicit confirmation gates before any action executes.

▶ 60s agent run — recorded demo

Placeholder — recording of the browser agent planning and executing a cross-tab task, with the confirmation gate visible.

✓ passed a 10-case adversarial security benchmark◦ self-reported · benchmark repo pending publication
Full case study → Chrome Web Store ↗ GitHub ↗

Problem

The network team ran compliance, delivery, and procurement workflows across manual processes and six separate enterprise systems, with no unified audit trail.

My role

Software Engineer Intern shipping features across all three platforms, including authentication and the core synchronization engine.

Constraints

Institutional security requirements: SSO, token rotation, per-route access control, and an audit log that had to be trustworthy rather than decorative.

What shipped

SAML SSO, rotating JWTs, CSRF protection, per-route RBAC, audit logging. A cancellable synchronization engine with an 8-worker pool, module-level locks and abort events, integrating 6 enterprise systems, validated against 167 backend tests✓ verified · source.

▶ architecture walkthrough

Placeholder — short recording over safe fixture data. Internal hostnames, tickets and operational details stay out.

✓ 167 backend tests passing across 6 integrated enterprise systems✓ verified · source
Full case study →

Problem

The paper Bells Test for spatial neglect is hard to standardize, time, and score consistently in clinic — and it captures nothing about how a patient searches space.

My role

Spatial Computing Developer & Undergraduate Researcher — built the Vision Pro application and the clinician-side control system; participated end-to-end in testing, analysis, and the write-up.

Constraints

Real patients in real clinical sessions. Reliability, timing accuracy, and a clinician controlling the session without touching the headset mattered more than polish.

What shipped

A Vision Pro test with 135 3D targets, gaze/pinch selection and automated timing, producing regional precision/recall and scan-path analysis. Local-Wi-Fi discovery, state sync, clinician iPad control and report transfer. Piloted with 30+ stroke patients at NUH and Alexandra Hospital~ preliminary · statistics from an early cohort; co-authored an ISMAR 2026 submission~ submitted · not yet accepted.

▶ interaction loop — synthetic data

Placeholder — clinician starts test → patient selects targets → scan path reviewed → report generated. Anonymized/synthetic fixtures only; no participant data.

~ ISMAR 2026 submission co-authored — under review, author list to be confirmed before publication claims
Full case study →

Problem

LLM-written company research is easy to produce and easy to distrust — a claim with no traceable source is worthless for due diligence.

My role

Designed and built the orchestration layer: the planner, the parallel agent pool, and the validation gate that decides what reaches the final report.

Constraints

A wrong but confident claim is worse than a missing one. The system had to prefer silence over fabrication.

What shipped

A planner dispatching financial, legal, news and risk agents in parallel, with claim-level provenance and fail-closed validation that suppresses unsupported findings before they reach the report◦ self-reported · repo available on request.

▶ traceable report — precomputed

Placeholder — a finished company report where every conclusion is clickable: claim → source → verification status. No paid credentials needed.

◦ provenance and validation behaviour demonstrable on request; public benchmark pending
Full case study → GitHub ↗

Problem

Rerankers applied to every QA prediction tend to fix some answers and break others — broad intervention is unstable.

My role

Fine-tuning, evaluation design, and the confidence-gating mechanism, in a CS4248 group project.

Constraints

A fixed benchmark (SQuAD) where regressions are visible immediately — you can't hide behind vibes.

What shipped

RoBERTa fine-tuned to 84.28 EM / 90.93 F1; the gate reranked only 55 of 10,570 low-margin predictions, lifting scores to 84.40 EM / 91.04 F1✓ verified · repo — every other prediction left untouched, deliberately.

✓ 84.40 EM / 91.04 F1 on SQuAD — reproducible from the repository✓ verified · repo
Full case study → GitHub ↗

Also on the bench: an M&A research pipeline built at Klimacap, a seven-crate asynchronous market-research platform in Rust, and a Whisper fine-tune for Singapore English. More on GitHub ↗

Experience

What I've been trusted with

  1. Jul 2026 — Aug 2026

    AI Engineer Intern · iEnergy Digital

    Deployed a retrieval-augmented interface over an incident knowledge base on AWS EC2 — embeddings, retrieval, and tool use returning source-grounded answers. Built read-only text-to-SQL for nested incident data, hybrid vector/BM25 retrieval, and a golden-set offline evaluation harness that converted clustered failures into targeted few-shot repairs.

  2. Jun 2026 — Aug 2026

    Undergraduate Teaching Assistant · National University of Singapore

    Taught 20+ students in NUS's introductory AI and machine learning module — search, logic, supervised learning, neural networks.

  3. Jan 2026 — Jun 2026

    Software Engineer Intern, Network Automation · NUS Information Technology

    Shipped features across 3 internal platforms; SSO, RBAC and audit logging; a cancellable 8-worker sync engine validated by 167 tests. Full case study ↑

  4. Apr 2025 — Jan 2026

    Spatial Computing Developer & Undergraduate Researcher · Interactive 3D Design Lab, NUS · with NUH

    Built the Vision Pro and iPad clinical assessment system; piloted with 30+ stroke patients; co-authored an ISMAR 2026 submission. Full case study ↑

    ran part-time alongside NUS IT — the overlap is intentional, not an error

  5. Sep 2025 — Dec 2025

    AI Developer Intern · Klimacap

    Built a multi-stage LLM research pipeline producing source-grounded M&A company and industry reports in DOCX with inline citations.

  6. May 2025 — Jul 2025

    Software Engineer Intern · Aurionpro Solutions

    Automated PL/SQL releases across Dev, SIT and UAT with reusable wrappers; shipped SmartLender Commercial UI features; prototyped extracting services from an Oracle/Java monolith.

  7. Apr 2025 — May 2025

    Lead Software Developer · Source Academy, NUS

    Maintained production CI/CD for TypeScript/React and Elixir services; diagnosed build failures, reviewed and merged contributions, and mentored developers through releases.

Beyond the code

The same job, off the clock

The through-line outside work is the one inside it: I end up in the position where someone has to be accountable for other people — a residential cluster, a tutorial room, a hackathon team.

Leading

Cluster Leader & Fire Warden, PGPR

2025 — present

Responsible for resident welfare and safety protocols across a diverse residential cluster at NUS — mediating disputes, coordinating community events, and being the person who answers when the alarm goes off.

Teaching

Volunteer Tutor, Teach SG (MOE)

2026

One-to-one tutoring for primary-school students in need — explaining ideas plainly and building learning confidence week over week.

Undergraduate TA, NUS

Jun — Aug 2026

Intro AI/ML module, 20+ students. Search, logic, supervised learning, neural networks.

Mentoring & community

Mentor, Apple Spatial Hack — AI hackathon

2026◦ details to confirm — event name, organiser, cohort

Mentored teams building spatial-computing and AI projects — the same stack as the clinical work, from the other side of the table.

Student community

Contributor to the 2025 NUSSU website◦ contribution scope to document and the NUS Mars Rover club site; mentored developers through Source Academy releases.

National Rank #1 — DXC Technology Coding Challenge◦ self-reported · year and verification link pending

Off the clock

The lab bench

Not case studies — things built because they were fun. They exist so the four above look selected, not exhaustive.

game · swift

Whack-A-MOLE

A reflex game in Swift — timing, hit detection, score state, kept deliberately small.

game

RPS, over-engineered

Rock-paper-scissors as an excuse to polish a real gameplay loop end to end. Web export planned.

discord bot · python

Music Bot

Queueing, playback control, and the usual audio-streaming headaches — built for one server of friends.

experiments

Smaller AI experiments

Forecasting trials, a Polymarket mirror bot, Spotify trend analysis — the bench where ideas get tested before they earn a case study.

Let's talk.

Best reached by email — I read everything.

arnav@u.nus.edu GitHub ↗ LinkedIn ↗

résumé for the current lens: download PDF ↓