Machine learning & software engineering

I'm a final-year computer science student. I work on machine learning and AI, from deep learning and reinforcement learning to LLMs, computer vision and AI safety, with research at IIT Kharagpur, Jadavpur University and NJIT.

On the side, I made GPT-2 in assembly, built the training pipeline for Kimi K3 from scratch, and contribute to open source.

Website · Resume · LinkedIn · Google Scholar · X

• IIT Kharagpur, May 2026 - present. Summer research intern with Dr. Sourangshu Bhattacharya. A multi-omic network for head-and-neck cancer, where DNA methylation alone tells oral from non-oral tumour origin at 0.937 accuracy.
• Jadavpur University, Nov 2025 - May 2026. Research intern at the CMATER Lab with Prof. Debotosh Bhattacharjee. A hybrid CNN-Transformer that segments breast thermograms at 0.9716 Dice.
• NJIT, Jun 2025 - Nov 2025. Summer research intern at the Learning-based Decision Making Lab with Dr. Arnob Ghosh. A semantic safety cost model for LLM alignment that agrees with human safety judgments 92% of the time.

• GPT-2 in Assembly: GPT-2 small in hand-written ARM64 assembly, with no C and no libc calls. It reads the original Hugging Face files, runs at about 100 tokens per second on one M1 core, and matches PyTorch token for token. Write-up on my blog.
• Kimi K3 for All: Moonshot released Kimi K3 with inference code only. I rebuilt the training pipeline from scratch, fixed the four places where the released code cannot train, and trained a 1.27B test model with it. Weights on Hugging Face, write-up on my blog.

• Recursive and Wrapper-Based Feature Selection for Breast Cancer Diagnosis and Prognosis. Ayushi Bhattacharjee, Arnesh Banerjee, Arpita Talukdar. 4th Analytics Global Conference (AGC 2026), Springer ASCAR Series. Oral presentation. 2026.
• Failure Modes of Large Language Models on Research-Level Mathematics: A Taxonomy and an Empirical Characterisation. Arnesh Banerjee, Ayushi Bhattacharjee. arXiv:2606.24902. 2026.
• Certifiable Safe RLHF: Semantic Grounding and Fixed Penalty Constraint Optimization for Safer LLM Alignment. Kartik Pandit, Sourav Ganguly, Arnesh Banerjee, Shaahin Angizi, Arnob Ghosh. arXiv:2510.03520. 2025.
• An Intelligent Weakly Supervised Framework for Breast Thermography Segmentation Using Hybrid CNN-Transformer Networks. Arnesh Banerjee, Debotosh Bhattacharjee. In preparation for Expert Systems with Applications.

• Zed Code Editor (9 merged PRs)
• Kubescape (8 merged PRs)
• OpenCV (5 merged PRs)
• Rust Language (2 merged PRs)
• Zen Browser (2 merged PRs)
• Activity Watch (2 merged PRs)
• Python (1 merged PR)
• Ghostty (1 merged PR)
• Root by CERN (1 merged PR)
• hnn-core by jonescompneurolab (1 merged PR)
• jenkins (1 merged PR)
• skopeo by podman-container-tools (1 merged PR)

• Let's make GPT-2 in assembly: GPT-2 small in ARM64 assembly, with no C and no libc calls. About 100 tokens per second on one M1 core, and the same output as PyTorch. November 2026.
• Next-Concept Modeling explained: A simple breakdown of a new method that generates text in layers, from a rough sketch to the final words, in a few parallel steps. November 2026.
• Amazon ML Challenge 2026 in a self-imposed 24 hours: Matching 10 million noisy business records to 1.7 million businesses across the US, India and an unseen France, for 0.988216 F0.5. September 2026.
• Kimi K3 for All: Rebuilding the training pipeline Moonshot never released, and the small model it produced along the way. August 2026.

Contact me: arneshbanerjee24 [at] gmail [dot] com