Skip to content
Aman SinghProjects

Selected projects

Systems I've built, shown working.

Case study · Slack Agent Builder Challenge 2026

Memora

A Slack agent for teams that forget why they decided things. It records the decision, the reason and the sources, and says how sure it is.

View project

Illustration · a sample thread

Illustration · rebuilt from a real session

KlarityCase study · EduX Hackathon 2026
Students prove what they understand by teaching an AI that gets things wrong on purpose. Here it plants a wrong idea about while loops, and the student catches it.MicrosoftEvent hostUniversity of CambridgeEvent partner

AI Showcase People's Choice, UTS Tech Fest 2026

LiPi Sort

Reads the handwritten postcode on a postcard and sends it to the right tray, for mailrooms too small for sorting machines. Built with two teammates.

YOLO11s · Synthetic data · EasyOCR · OpenCV · Streamlit

Try the app ↗Read the report →

Interactive illustration · the app's steps, redrawn

01 / 07 · scripted run of the real pipeline

Photo

Cards arrive the way people photograph them: tilted, in perspective, sometimes upside down.

1 phone photo

How it works, with photos and the demo video

The app straightens the card, turns it upright, finds 10 regions with a fine-tuned YOLO11s, reads the postcode with EasyOCR (with a vision-model fallback when it isn't sure) and looks the suburb up on Zippopotam.us. Then a robotic arm we built puts the card in its tray. We trained on 1,310 postcards, 1,200 of them fake, because real mail is private.

Recording · plays on YouTube

Research and seismic engine

South Asia Earthquake Severity

A random forest that predicts how severe an earthquake in South Asia will be, trained on USGS records. The C++ engine and PyTorch friction layers I built alongside it have their own write-up.

Random forest · PyTorch · C++ · USGS data

View source ↗

How much energy?

Interactive illustration

PS
7.2 Mw

951 kilotonnes of TNT

Roughly 1,995× the energy of a magnitude 5. Each whole step is about 32× more energy. The trace is drawn on a log scale, or a magnitude 4 would be a flat line.

Resume parser and keyword engine

ReCV

Rewrites a resume so applicant tracking systems can read it, then checks it against the keywords in the job ad.

Next.js · Groq · React PDF · TypeScript

Try the keyword check

Interactive illustration · a simplified version of the check

5 of 7 keywords from the job ad found

Missing: APIs, AWS. The real app also rewrites the resume with an LLM.

UTS Cybersecurity Tech Fest 2026

Canvas Incident Response Plan

A blue-team retrospective of the 2026 ShinyHunters breach of Canvas LMS. Reuters reported 8,809 institutions affected; the attackers claim data on about 275 million users. We mapped every step to MITRE ATT&CK and rebuilt the response under the ACSC framework.

MITRE ATT&CK v19 · ACSC IR framework · Threat intel · Zero trust

Read the report →
Step through the breach

Interactive illustration · our reconstruction, not confirmed detail

Stored XSS in a free-tier account ran JavaScript in an admin session. Step through the chain to watch it spread.

We reconstructed the attack timeline from stored XSS to ransom note, then rewrote the response plan against the ACSC framework and Australia's breach reporting obligations. Built with two teammates.

Smaller things I built for fun, like a ray caster and a matching engine, live in the Workshop.