Technical Mentor @ Monash University

I build AI features that keep working when the model doesn't.

Right now I am building RAY/OS, my own AI operating system. Before that I shipped NeighbourFit, the Monash expo winner. I also mentor student software teams at Monash, where I did my Master of AI.

Break the modelSee the work

Melbourne --:--building something

Go on, try to break my model. It is allowed.

The demo

Break the model.

Picture someone searching for a suburb at 11pm just as the AI provider goes down. If the page shows an error, they leave and do not come back. If it still hands them a ranked answer, they never notice, and they keep trusting the product.

That is how I built NeighbourFit: a deterministic scoring engine always answers first, and the language model only ever improves the answer. Pick a failure and watch what the user would see.

Simulation - fictional suburbs, no real model called

Pick a button to start

You: student, no car, lives for cafes

The scoring engine answers first. The model only ever improves the answer. Try to make it fail.

Your preferences
Scoring engine
LLM re-rankLlama 3.3 (simulated)
Schema check
Result card

Selected work

Seven things I built.

A team product, a solo app, systems built from the maths up, the system I run my life on, and three smaller builds you can play with in your browser. Poke the pins, knock out a provider - the cards are live.

Expo winner, Monash 2026

NeighbourFit

Choosing where to live in Melbourne meant cross-referencing transport, amenities and demographics across separate sites, with nothing weighing it against your own priorities.

How it works

A suburb recommender where a deterministic scoring engine always answers first and Llama 3.3 only re-ranks, summarises and reads voice input. Enforced JSON schemas and graceful fallbacks mean the product keeps working when the model does not.

290+Melbourne suburbs scored
83%persona accuracy on voice input
  • Vue 3
  • Flask
  • AWS Lambda
  • PostgreSQL
  • Llama 3.3 70B
Read the case study

Solo, shipped

Outfit Picker

A bring-your-own-key AI app breaks the moment one provider returns something unexpected.

How it works

A local-first PWA where OpenAI, Anthropic, Gemini and Groq swap at runtime behind one interface, each with a four-tier fallback ladder for structured output. Photos and keys never leave the device.

4AI providers behind one interface
0servers that see your data
  • React
  • Vite
  • IndexedDB
  • PWA
Live demoCode

From first principles

AI Systems

Using a library is easy. Knowing what it does when it breaks is the job.

How it works

Four DQN agents trained from scratch with replay buffers and soft target updates. Optimisers written by hand on raw tensors. A Pacman team that tracks hidden opponents with a particle filter and plans with risk-aware A*.

4cooperating agents, trained from scratch
3optimisers written by hand
  • Python
  • NumPy
  • PyTorch
Code

Personal, running daily

RAY/OS

My life was spread across mail, calendars, task lists and endless chat threads, and none of it remembered anything.

How it works

A second brain on Claude and an Obsidian vault: a pipeline for notes, skills that encode how I work, agents that run on a schedule, and a voice console I can reach from my phone. It drafts, I decide.

Dailythe system I actually run my week on
Humanin the loop for anything that leaves it
  • Claude
  • Obsidian
  • Python
  • Flask
  • MCP
See how it works

Also built, try them in your browser

Python, XGBoost

Melbourne house prices

XGBoost with an honest 80% price range, running entirely in your browser. 10.3% median error on sales it never saw.

Play with itCode

Python, TF-IDF

Movie recommender

Pick a film, get five that share its cast, director, genre or plot, with the reason shown for each.

Play with itCode

Python, Pygame

Snake

Pygame compiled to WebAssembly, so the desktop game plays in the browser. Swipe on a phone.

Play with itCode

Everything fails, all the time.
Werner Vogels, CTO of Amazon

Hi, I'm Rishabh. I build AI features, ship them, and design for the day the model fails.

Focused on applied AI: LLM products, agents, and the deterministic systems that keep them honest.

Computer science in Delhi, with a camera never far away - I ran the college film and photography society. Then brand operations at a Delhi design studio, where I automated order tracking across Shopify, Stripe and Mailchimp and got a day a week back for the team. Then a Master of AI at Monash, and a turn toward multi-agent systems, planning and LLM products that have to hold up for real users - like NeighbourFit, which won the Expo.

Origin trace
  1. Delhi, computer science
  2. Design and brand ops
  3. Master of AI, Monash
  4. AI products
  • Master of AI, Monash
  • Expo award winner
  • LLMs and agents
  • Photographer

78

students mentored across 13 teams

290+

suburbs scored in NeighbourFit

3

LLM features shipped to a live product

7,500

units delivered for Mini Cooper in one month

My stack

What I work with.

The core of it. Hover or tap a logo to see where I used it, then jump straight to that project.

AI / LLM

ML

Backend

Frontend

Design

Track record

Where I have been.

  1. 2026 - nowMonash UniversitySessional Teaching Associate, Technical Mentor FIT3047AI-usage standards for a whole unit

    Mentor 13 student teams through an industry-sponsored software project. Set the unit's standards for AI-assisted development: which assistants students may use, where they may operate in a codebase, and how their work is logged and reviewed. Run product-owner sessions that translate technical progress for non-technical clients.

  2. 2024 - 2026Monash UniversityMaster of Artificial IntelligenceExpo award winner

    Deep learning, machine learning, multi-agent systems, planning and automated reasoning, discrete optimisation. Capstone team won the PG Industry Experience Expo.

  3. 2023 - 2024Say It With A PinBrand Operations Manager7,500 units for Mini Cooper in a month

    Built Zapier and n8n pipelines across Shopify, Stripe and Mailchimp that replaced manual status checks and saved a day or more of coordination every week. Ran B2B accounts including BMW and Third Wave Coffee, and moved the studio onto one shared workflow.

  4. 2022Anev MediaWebsite Development Intern+40% organic traffic

    Rebuilt the site on WordPress and shipped features in Python, JavaScript and Firebase, with analytics-driven SEO work alongside.

How I work

RAY/OSMy personal AI operating system.

A second brain I built on Claude and an Obsidian vault. Things flow in, get sorted into a pipeline, and a set of skills and scheduled agents keep it honest. I talk to it through a console on my laptop or phone. It drafts. I decide.

Sources

Mail, calendar, tasks and years of chat history flow in. Nothing is trusted just because it arrived - everything lands in an inbox first.

ray/os console - example, fictional data

Pick a question below.

Principles

Things I learned the hard way.

Four lessons from shipping NeighbourFit, written down so I stop relearning them.

Click a note for the real story behind it.

Outside the code

Before the models, a camera.

I ran Zephyr, my college film and photography society, led a 20-person coverage team at TEDxCVS 5.0, and freelanced in photography and design. It is still how I think about interfaces: frame it, cut what does not belong, and make the one thing that matters impossible to miss.

Now

What I'm up to.

Updated September 2026

Building
RAY/OS, my personal AI operating system - and this site.
Mentoring
Thirteen student software teams through an industry project at Monash.
Open to
Graduate and junior AI and ML engineering roles.
Weekends
Out with a camera, when the week allows.

Contact

Let's build something
that doesn't break.

LinkedInGitHubDownload CV

Melbourne, Australia. Full work rights to August 2029, no sponsorship required.