Brisbane, Australia · Open to relocation · Full work rights (Subclass 485)
Hi, I'm Akshat.I build
Software engineer with a Master of Information Technology from The University of Queensland. I build full-stack and cloud applications, from React interfaces to Python and Node backends, and I like turning a rough idea into something people can actually use.
Drag, swipe, use the arrow keys or the buttons. The first one is the one I'd start with.
01 / 10
Featured · Live Chrome extension01
ViewSeen
A Chrome extension for sending what you find online to friends and talking about it right there. I built it end to end and published it on the Chrome Web Store, where anyone can install it.
Right-click anything
Share card
Friend's chat
Opens at that exact spot
Right-click a post, comment, video, image or selected text to share it with friends or any other app. Shared text opens on the other side scrolled to and highlighting those exact words, using URL text fragments.
Messages are end-to-end encrypted: each pair of users derives a shared key with ECDH P-256 and HKDF, and every message is sealed with AES-GCM and a fresh IV. Trade-off I chose on purpose: private keys live with the account so history works on every device, which means it protects against other users and data leaks, not against the project operator.
Real-time chat on Firestore with replies, reactions, GIFs, typing indicators, read receipts and unsend. Message paging keeps every loaded page live, so an unsend always shows up and a message being sent never jumps out of order.
Asks for no access to the websites you visit. It reads a page only when you choose to share from it, and an offscreen document watches for new messages to drive notifications and the toolbar badge.
Personal invite links send a friend request in one tap. Firestore security rules guard friendships and messages.
A Chrome extension that clicks through the university portal for you. Ask for something in plain English and it finds the page. Built with a team at the UQCS hackathon 2025.
Chat in extension
Local Python server
Vector search
Step plan
Playwright clicks
The front end is a React and Vite Chrome extension (Manifest V3) with a chat window. Chat history is stored in Firebase.
A small Python server turns the request into steps. It converts the sentence into an embedding and finds the closest stored navigation example in a ChromaDB vector database.
Playwright then drives the user's own open Chrome tab through those steps: click, hover, fill, type, select and wait.
Portals are awkward, so it waits for the Angular page to settle, matches button labels with fuzzy text matching, caches the selectors it has found, and follows new tabs.
Early prototypes asked a local Llama model (through Ollama) to plan the steps directly.
A smart parking platform where users find a space on a map, book it and manage their bookings. Built twice on purpose: as one server and as three services.
React + Leaflet
↓
auth
booking
parking
one SQLite database per service
Monolith (Express and Sequelize) and microservices versions side by side, so the two designs can be compared.
JWT sessions, bcrypt password hashing, Dockerfile per service, Docker Compose for local runs.
Kubernetes manifests with ingress routing and a Horizontal Pod Autoscaler, deployed to Google Kubernetes Engine.
Load testing and tuning improved the booking workflow by about 40%.
Restaurant ordering from the table. Guests scan a QR code, order from their phone, and the kitchen sees it on a dashboard. It is deployed and you can open it.
Scan table QR
Browse and order
Kitchen dashboard
Received → Completed
Each table has its own link, so an order arrives already knowing the restaurant and table number. Guests do not need an account.
Staff build menus, categories and items with photos and prices, watch incoming orders, move them from Received to Processing to Completed, see which tables are occupied, and print a QR code for each table.
Django with a custom user model tied to a restaurant. Staff endpoints sit behind login, guest endpoints are public, and the data is modelled as restaurant, menu, category, item, order and order item.
Hosted on Render with a Neon Postgres database. I moved the secret key and database credentials out of the code into environment variables.
Listens to a tabletop D&D session, writes down what was said, and lets you ask the campaign archive questions afterwards. Everything runs on your own machine.
MicWhisperOllama LLMArchive
Real-time transcription with Faster-Whisper and voice activity detection, then a local Gemma model picks out events such as combat, item transfers and quest updates.
Falls back to rule-based pattern matching when no model is available.
Three backups of each session: SQLite, JSON summaries and raw transcript logs.
FastAPI backend, browser front end, character sheets, inventory and journal. Team project: I built modular components and documented the architecture.
A small programming language written from scratch, plus a bug-finder that combines random mutation with a constraint solver.
# pseudocode: why random mutation strugglesif x == 7331:
y = 100 / (x - 7331) # divide by zero
Illustration of the kind of bug it targets. Random mutation almost never guesses 7331; the solver derives it.
Lexer, recursive-descent parser, syntax tree and a tree-walking interpreter that records branch and statement coverage.
The random fuzzer mutates inputs and keeps only those that reach new code.
When random search stalls, Z3 symbolic execution collects the conditions along a path and solves them to produce an input that reaches the unexplored branch.
Finds division by zero and assertion failures guarded by exact-match conditions that random mutation almost never hits.
Three university assignments that build from path finding to planning under uncertainty to deep reinforcement learning.
1A* with a custom admissible heuristic
2Value iteration, policy iteration, Q-learning
3Deep Q-Networks in PyTorch
Cheese Hunter grid world: a heuristic built from precomputed lever-to-goal distances expands far fewer nodes than uniform cost search.
Planning and learning methods converge to the same optimal policy.
DQN on CartPole and LunarLander, with experiments on learning rate, epsilon decay, target-network updates and network size. The LunarLander agent averaged about 210 reward over 100 episodes, above the 200 success threshold.
Software Engineer - AI Trainer DataAnnotation, freelance
Evaluate AI-generated answers on coding and reasoning tasks against detailed rubrics, write coding prompts that stress-test frontier models, and write test cases that expose edge-case failures in open-source Python codebases.
Master of Information Technology The University of Queensland
GPA 6 out of 7. Machine learning, artificial intelligence, cloud computing and advanced software engineering.
Learned Java, Cassandra and the Quartz scheduler quickly enough to deliver a working Hacker News clone with scheduled background jobs inside the programme. Received a full-time offer and declined it to study at UQ.
Built an Android app in Kotlin, cut local query time by 10 to 15% by optimising SQLite queries, and raised a barcode scan success rate from about 90% to over 95% after tracing the root cause. Integrated Twilio SMS notifications and answered more than 50 developer support tickets.
B.Tech, Electronics and Communication Puducherry Technological University
CGPA 8.58 out of 10.
Problem solving
Competitive programming
I practise data structures and algorithms on LeetCode, and I take part in its timed weekly contests.
0problems solved
0contest rating
0contests entered
0of rated users globally
0LeetCode profile rank
Best finish: rank 809 out of more than 20,000 in Weekly Contest 372.