I’m Mrudula, an AI Engineering student at Amrita Vishwa Vidyapeetham.

I keep ending up in projects where a model has to survive contact with something annoying: two badly synchronized cameras, noisy biosignals, adversarial molecular strings, incomplete financial evidence, people speaking into microphones, or a UI that has to make all of that legible.

Right now the biggest rabbit hole is markerless motion capture for gait analysis: stereo reconstruction, pose estimation, temporal cleanup, kinematics, gait events, quality gates, and the much less glamorous question of whether any number coming out of the pipeline deserves to be trusted.

I also build products because sometimes the research question needs an interface before it needs another notebook.

vision motion + gait
signals EEG
security smart grids
robustness adversarial ML
systems evaluation + tooling
product interfaces + apps

Coimbatore, India
B.Tech · AI Engineering
2024–2028

research ↔ software ↔ product

A dual-camera markerless motion-capture system for 3D human pose and gait analysis: calibration, synchronization, triangulation, temporal stabilization, kinematics, gait metrics, live quality checks, and validation tooling.

Python OpenCV MediaPipe 3D geometry gait

_{Research build · repository currently private while validation work is in progress.}

Adversarial robustness experiments for materials sequence modelling: chemically valid perturbations, attack search, model-shortcut analysis, adversarial training, and attack/defense evaluation.

PyTorch RDKit Transformers MCMC robustness

A mobile-first exploration / habit product with server-authoritative progression, rarity and grading systems, verification flows, append-only coin accounting, and a deliberately strange little forest wrapped around the whole thing.

React Vite Supabase Express Capacitor

Evidence-first financial reconciliation using graph structure, staged matching, exception handling, provenance checks, and an AI layer that is allowed to explain evidence but not manufacture accounting truth.

Python FastAPI NetworkX evaluation

ML-based intrusion detection for IEC 60870-5-104 traffic. This work became my first-author IEEE Access paper.

Python scikit-learn cybersecurity evaluation

A voice-first language-learning experiment with conversational tutors, translation, memory, and learning analytics.

Next.js TypeScript Supabase voice UX

Smart Grid Intrusion Detection for IEC 60870-5-104
IEEE Access · 2025 · first author
paper ↗

PV Fault Detection with Pyramid Network + GRU
Energy Conversion and Management: X · 2025

A lot of my projects eventually collapse into the same annoying sentence:

what would make this result believable?

Sometimes the answer is a benchmark. Sometimes a quality gate. Sometimes an ablation, a failure case, provenance, or admitting the system does not know yet.

That part is usually more interesting to me than squeezing another decimal out of the headline metric.