Speaker
Description
Every software developer uses Git to track code changes line-by-line as a project converges toward a stable release. But when data scientists and ML researchers try to use Git to manage exploratory experiments, the workflow breaks down.
That is because software engineering and computational research have opposite goals. Engineering moves toward convergence (building one deterministic codebase). Research requires constant divergence—testing dozens of model architectures, dataset splits, loss functions, and parameter combinations in parallel.
When you force Git to track experiment variations, your commit log gets polluted with messages like "try lr 0.001", your Jupyter notebooks refuse to merge cleanly, and your folder quickly fills up with exp1_v2_final_REAL.ipynb. Meanwhile, production MLOps tools log run metrics after the fact, but they don't help you structure or query your exploratory R&D workflow on your local machine.
In this 20-minute talk, we look at why Git's line-by-line commit tree isn't built for experiment search spaces, and how adding a lightweight open-source component layer in Python (using PyLabFlow and SQLite) lets you keep Git for your source code while keeping your experiment variations structured, deduplicated, and queryable—100% offline.
Session author's bio
I am a Research Systems Engineer and the Founder of ExperQuick Research Infra, where I build domain-independent infrastructure for computational research. My work focuses on standardizing experiment management and reproducible workflows across AI, mathematics, computational biology, simulation, and other research domains.
I conceptualized and developed PyLabFlow (formerly PyTorchLabFlow), an open-source framework for structured experiment management that has surpassed 25,000 downloads and supported published research in the Journal of Applied Bioanalysis. My academic background includes a Master's in Data Science from the Defence Institute of Advanced Technology (DIAT–DRDO). As an advocate for reproducible science and local-first open-source research infrastructure, I regularly speak at industry events, including UbuCon India, Open Source Summit India, and AI in The New Era.
Any other info we should know?
What attendees will get out of it:
- A clear explanation of why Git struggles with experiment variations and notebook workflows.
- Practical insights on separating code versioning (Git) from experiment state management.
- A look at how to run a local-first, privacy-friendly experiment workflow on Ubuntu using Python and SQLite without needing cloud tracking platforms.
Format: Technical presentation with a brief demo and Q&A.
| Agree to Privacy Policy and Notice | I agree |
|---|---|
| Please confirm that there are included headshots of all speakers in their profiles | Yes |
| Social Media | X (Twitter): https://x.com/BBEK_Anand GitHub: https://github.com/BBEK-Anand LinkedIn: https://linkedin.com/in/bbek-anand |
| In Person Attendance | In-person |
| Level of Difficulty | Intermediate |