S·K·C

Portfolio · 2026 ML researcher and engineer.

Shushant Kumar Choudhary

Portrait of Shushant Kumar Choudhary

I do machine-learning research on quantum error correction, world-model control and AI-agent safety. I build decoders, simulators, agents and small speech models, and I test each one carefully.

Question. Build. Measure.

Project brief · No. 001SurgeryNet

Can a graph neural decoder trained on memory experiments also work during lattice surgery?

Metric
logical error rate vs MWPM
Seeds
8 × 30,000 shots
Result
transfers to surgery

Every project comes with tests and clear documentation. Here are five selected projects.

Project 01 · quantum error correction

SurgeryNet

The first reported learned-decoder error rates for lattice-surgery merge and split. MLP and CNN decoders can't transfer to surgery; a graph neural network can.

PyTorch Geometric · Stim · PyMatching · 9 experiments · 84 passing tests · LaTeX paper draft

Read the code →

Project 02 · drone-swarm security · Terrier Cyber Quest 2026

VISHWAS

One drone lies about another. Vote-only consensus believes it. ClaimCheck audits each accusation against the swarm's broadcast ranging data.

92honest drones wrongly removed
vote-only consensus
0honest drones wrongly removed
ClaimCheck, 538 missions

Attacker catch rate 32% → 50%. Code →

Project 03 · physics-structured world models

Hamiltonian World Models

A port-Hamiltonian world model against MLP, PINN, latent neural ODE and a Dreamer-style RSSM. 4 physics systems, 60 runs, bootstrap confidence intervals.

energy drift time →
Well-fit seeds cut drift from 0.15–0.27 to 0.002–0.006. Out-of-distribution seeds are analysed in the paper.

Benchmarked against four baselines, with a full write-up on where structure helps. Paper draft →

Project 04 · AI-agent safety

Mirage

An adversarial web for measuring how browser agents fall for prompt injection. Hidden payloads sit in snapshotted cloud sandboxes and are scored server-side.

  • 3 runs per cell, isolation control at 1.00 agreement
  • Found and reproduced 2 bugs in the live Solari API, shipped workarounds
Read the code →

Project 05 · speech

Aaवाz

Text-to-speech for 12 Indian languages in 2.4M parameters. Distilled from Piper/VITS teachers into a 4.8 MB model that runs on a CPU with only NumPy at 0.09× real time.

Press a language. These are the model's own samples.

Read the code →

More projects

    Evidence

    Field work

    • Indian Army Terrier Cyber Quest 2026, AI Kavach track
    • Smart India Hackathon 2026, Team Orbit (SIH26078 · SIH26054)
    • Hlyn AI red-teaming challenge 2026
    • Kaggle: Kagriculture, K-12 Student Misconception
    • Ant Media Hackathon · Algo Olympia · ByteBuilt

    Training

    • B.E. Computer Science (AI & ML), Chandigarh University, 2025–2029
    • Python · PyTorch · PyG · PennyLane · Stim · TypeScript · React · FastAPI · Kotlin · C++

    Let’s work together.

    Open to research collaborations, internships and ML engineering roles.

    Mohali, Punjab · © 2026 Shushant Kumar Choudhary