- I'm an ML Research Engineer at Animatica, working on generative models for 3D character animation.
- My research is in mechanistic interpretability and AI safety, with side interests in reinforcement learning and multimodal AI.
- I recently completed my B.Eng. in Data Engineering (thesis graded 5/5, graduated early 2026).
A Mechanistic Study of AI-Text Detection Neurons in Frozen BERT: Sparse Probing and Activation Patching on RAID Accepted to EMNLP 2026, main conference (Interpretability and Analysis of Models for NLP).
We identify small, stable sets of neurons (45-62 per generator) in frozen BERT that are causally sufficient for AI-text detection, using L1/L2 sparse probing and bidirectional activation patching on the RAID benchmark. Code: Gradient-PG/ai-text-detection-neurons
- Motion Generation → Training pipelines and models for motion inbetweening from character rigs at Animatica.
- Mechanistic Interpretability → Following up on the EMNLP work: how detection-relevant features transfer across generators and domains, and what the neurons actually encode.
- TRACE (RL Explainability) → Shapley-value attribution (SVERL) for deep RL agents, attributing return to temporal phases of behaviour and extracting readable structure from learned policies. LunarLander and CarRacing, with Gradient PG.
- autoresearcher → Autonomous research agent for ML codebases. Define directions, leave it running, and it explores the code, runs experiments, reads papers and iterates overnight.
- warehouse-bot-training → Custom PPO with multimodal observations (CNN plus task embeddings) for a warehouse navigation agent in Unity.
- Gradient PG: took over the club when it was down to four people and inactive, rebuilt it to 40+ members, and served as president for a year. Set up a research track targeting conference publications, alongside workshops, meetups and competitions. Now a senior member advising on research direction and methodology.
- Previously: ML Engineer at TeaCode.io, R&D Engineer at Dataedo, full-stack developer at Sidnet.
- Programming: Python (PyTorch, TensorFlow, Keras), JavaScript/TypeScript (React, React Native, Node.js), C# (.NET/Unity), C/C++
- AI/ML: Deep Learning, Mechanistic Interpretability, Natural Language Processing, Reinforcement Learning, Computer Vision, LLMs, Generative Models for Motion
- Other: Data pipelines, experiment tooling, web & mobile development, EDA
If you're looking for a collaborator on interpretability, AI safety, or applied ML research, feel free to reach out.