Abhinaba Chakraborty

My daily drivers are an Apple MacBook M1 for personal projects and a Dell Latitude 5540 (Intel i7) for university work. On the Dell, I run Arch Linux with the i3 tiling manager — I really like the minimal overhead and customizability of i3.

For heavy research workloads, I lean on imec's IDLab testbeds. I use VirtualWall (over 550 bare-metal servers) to spin up networking and distributed systems experiments. For training models, I use GPULab for NVIDIA V100s and RTX GPUs. For edge AI inference, my go-to is the NVIDIA Jetson Orin Nano — about 40 TOPS, perfect for profiling tensor ops locally.

I rely heavily on the Antigravity CLI to run a multi-agent workflow. My main engine for planning architecture and writing code is Gemini 3.1 Pro. After it writes the code, I pass the results to Claude for a harsh critique to catch edge cases and logic bugs. Claude tokens can get expensive fast, so I have to be careful with this loop — but feeding Claude's corrections back into Gemini works surprisingly well.

For monitoring long-running remote training jobs, I spin up cheap, lightweight Gemini Flash agents that watch logs over SSH and handle basic errors, saving a lot of manual babysitting.

For most things — large Python projects, Jupyter, Docker, remote SSH to my Jetson cluster — I live in VS Code. For a quick config edit or remote debug, I fire up Neovim (native LSP).

Indispensable extensions: Python + Pylance, C/C++, Remote-SSH, Docker, GitLens, Jupyter. For ONNX models, Netron.

I use Alacritty for the GPU-accelerated snappiness, tmux to keep sessions alive, and a few shell scripts that set up multi-pane layouts for parallel experiments.

Shell is Zsh with Oh My Zsh and Powerlevel10k. Unlimited history plus Ctrl+R and fzf for finding a three-month-old command is a lifesaver.

PyTorch for training and prototyping, ONNX Runtime for portable inference, TensorRT for optimized Jetson deployment. Most models convert to ONNX before hitting the edge cluster — it makes profiling and benchmarking far more reproducible.

For architecture and network simulation: SST Core / Elements and NS-3. Containerization via Docker and K3s. Analysis in NumPy, Pandas, Matplotlib and Seaborn.

Overleaf exclusively for paper writing — real-time collaboration with co-authors is indispensable. References via Zotero; diagrams via draw.io.

Notes and project management in Notion; day-to-day on Slack; code in GitLab (UGent) or GitHub (personal).

Git for everything. For research projects with multiple experiments running at once, git worktrees are invaluable — each branch gets its own working directory without duplicating the repo. Commit messages stay descriptive but simple; the experiment logs are what matter for reproducibility.

• Netron — visualizer for neural networks and ML models.
• KWOK — Kubernetes without kubelet; simulate clusters at scale.
• MathJax — beautiful math in browsers.
• The first website — the original WWW project page at CERN.
• EU Swarm Workshop — European research on edge intelligence swarms.