“One must imagine Sisyphus happy.” ― Albert Camus

Open Source AI

• sktime (ex-Senior AI Engineer & Maintainer): Redesigned the Deep Learning model architecture, built a PyTorch-based RNN classifier/regressor for time-series, and led the TensorFlow-to-PyTorch backend migration. Also redesigned the Forecasting Horizon module to decouple it from pandas, and contributed to dependency resolution, testing infrastructure, and reviewing community PRs across the framework.
• LightlySSL: Implementing LeJEPA (self-supervised learning for vision foundation models by Balestriero & LeCun et al.) - core modules, loss functions, multi-view transforms, and distributed training. Migrated the project's linting and formatting infrastructure to ruff.

Exploring AI Safety ☢️

• Replicating Anthropic's Synthetic Document Finetuning study to deepen my understanding of how post-training interventions affect model behaviour, on Llama-3.1-8B using gpt-4o-mini generation and QLoRA finetuning on free Kaggle GPUs). Check here.
• Contributing to Inspect AI and TransformerLens, both crucial pieces of AI safety tooling and infrastructure (LLM evaluation and mechanistic interpretability).
• Completed Blue Dot Impact's Technical AI Safety course (alignment, RLHF, mechanistic interpretability, evaluations, red-teaming, scalable oversight)
• I'm open to collaborations in AI Safety, feel free to reach out via LinkedIn mentioned at the end.

• ML Tech Lead at Honeywell Aerospace: architected an org-wide ML framework, shipped models across anomaly detection, NLP search, and time-series prediction ✈️
• Built Bayesian probabilistic forecasting at PriceLabs, serving 500K+ users daily 📈
• Co-founded Zencat Studios, a product and UX design agency for AI products 📲
• Built a wooden cabin in the lower Himalayas 🪵

• going for long walks 🚶
• reading and all things literary 📚
• looking at trees outside my window 🌲

• How interpretability could help bridge the gap between what a model knows and what we can verify it knows.
• Why classical statistical models worked well with less data, while modern ML models built on the same foundations are so data-hungry.
• The boundary between scalable oversight and human judgement, and how to design systems where the former supports the latter rather than replacing it.

LinkedIn: https://www.linkedin.com/in/nirbhai/

👨‍🏫 Work-related 📒

• Designing Machine Learning Systems - Chip Huyen
• Concrete Problems in AI Safety (paper) - Amodei et al.
• Human Compatible: AI & the problem of control - Stuart Russell

🤩 Just for fun 📚

• A Chess Story - Stefan Zweig
• Artificial Condition (The Murderbot Diaries, #2) - Martha Wells
• my wife's thoughts (trying to) 💭

⚡ Fun fact: import antigravity