# Arsh Kumar — MLE + SWE

**The bigger full-stack: from model building to the software on top.**

Machine learning engineer and software engineer. BEng Computer Science (First Class), University of Edinburgh, 2026.

I can train models and also steer agentic software development. My education and experience have taught me how to take software decisions *and* be able to debug PyTorch.

- Website: https://arshkumar.com
- Email: hello@arshkumar.com
- LinkedIn: https://www.linkedin.com/in/kumararsh/
- GitHub: https://github.com/thearshkumar

The stack I work across, bottom to top: 01 Data → 02 Model → 03 Train → 04 Serve → 05 Product.

## 01 — Selected work

Four things I built, tested or wrote up.

### Research paper search (2026)

Hybrid search over arXiv papers by conceptual similarity to a query. I added a "core idea" per paper, which retrieved better than abstract embeddings: **+21% vs abstract embeddings**.

- Tags: software, model, Supabase, Ollama, Vercel
- Layers: Data, Model, Serve, Product
- Try it live: https://search.arshkumar.com

### Palm-vein biometrics (2025)

On Nethermind's hardware ML team I initiated synthetic image generation with a diffusion model and built the pipeline, from pre-processing to inference, for it. Better dataset filtering moved a business metric in week one.

- Tags: model, diffusion, ViT, PyTorch
- Layers: Data, Model, Train, Serve
- Nethermind · Sep–Dec 2025

### Forecasting benchmark (2025)

Wrapped three forecasting packages in a small OOP layer so training and inference take a few lines. Tested GPU use and automated workflows; handled data pre-processing using pandas.

- Tags: software, Darts, pandas, PyTorch
- Layers: Model, Train, Serve
- With Prof. Jeremy Bejarano, UChicago
- Working paper: https://www.financialresearch.gov/working-papers/2026/08/25/time-series-forecasting-methods-financial-markets/
- OFR blog: https://www.financialresearch.gov/the-ofr-blog/2026/08/25/deep-learning-methods-improve-financial-forecasts/
- GitHub repo: https://github.com/jmbejara/ftsfr

### AI and nuclear command and control (2025)

A 14-page paper on the explainability, robustness and safety limits of deep models in high-stakes nuclear systems. Accepted to the 2025 SPSA summer conference.

- Tags: research, AI safety, NC3
- Layers: Model
- XLab Nuclear Risk Fellow

## 02 — Projects

What doesn't fit on a CV, and the years behind it.

### Personal: building for my dad's business (~9 years, ongoing)

I've been helping my dad with his business for around nine years, mostly on the digital side. I made designs on Canva and ran Meta ads. Before agents existed I built WordPress sites with Elementor, and a Shopify store that wasn't custom code but was entirely hand-made from templates with the Shopify builder. This year I built three different websites for him using agentic development.

It was never only building. I also sat in on decisions and planning, and the business's needs set what I had to learn next, so I picked up skills on the go. That's why AI feels natural to me: stacks change, but the principles and the decisions behind them mostly stay the same.

Languages show it. We now lean on type-safe languages, and my C++ and Java make static typing comfortable. I've also used Python heavily for ML, so I know what duck typing makes possible. The same goes for OOP and design patterns: knowing when they help is what lets me steer coding agents well.

I started early with LLMs too. When they were still new, I connected ChatGPT to WhatsApp before it was an official feature. I also chained speech-to-text and text-to-speech models to reply in my own voice. Along the way I learned a lot from Andrej Karpathy's online lecture series.

- Tags: software, Canva, Meta Ads, WordPress, Elementor, Shopify, agentic dev, WhatsApp bot

Try it live:

- **ResearchRAG search** — https://search.arshkumar.com — find arXiv papers by the idea you have in mind, not the keywords.
- **Latent Lab** — https://play.arshkumar.com — give a neural network two numbers and it draws a doodle. A VAE trained on 240,000 Quick, Draw! doodles, running in the browser.

## 03 — Stack

Data and training on one side, serving on the other.

I started Computer Science due to my passion for coding. During university, I started going deeper into ML. Inadvertently, the two gave me a handle on the larger full stack in the world of AI.

| Area | Skills |
| --- | --- |
| Languages | Python, C, C++, Java, Haskell |
| Models | Transformer, ViT, Diffusion, VAE, GAN, ConvNet, RNN, LSTM, GRU, LLM |
| Libraries | PyTorch, NumPy, pandas |
| Shipping | Vercel, Supabase, Ollama, Node.js, Git |
| Tools | Claude Code, Lambda AI, gdb, VS Code |

## 04 — CV

Experience and education (updated 2026-10).

### Intern — Nethermind, Edinburgh (Sep–Dec 2025)

Hardware division ML team, working on palm-vein biometric verification. Synthetic data with diffusion models, ViT training and evaluation, GPU-offloaded embedding search, and GAN-based pen testing.

### Research Collaborator — Prof. Jeremy Bejarano, UChicago (Jun–Sep 2025)

Financial time-series forecasting benchmark. Built the model classes, GPU and workflow tests, and the data pre-processing pipeline.

- Working paper: https://www.financialresearch.gov/working-papers/2026/08/25/time-series-forecasting-methods-financial-markets/
- OFR blog: https://www.financialresearch.gov/the-ofr-blog/2026/08/25/deep-learning-methods-improve-financial-forecasts/
- GitHub repo: https://github.com/jmbejara/ftsfr

### XLab Nuclear Risk Fellow — University of Chicago (Jan–Mar 2025)

Wrote a 14-page paper on AI in nuclear command, control and communications, accepted to the 2025 SPSA summer conference. Also an XLab AI Safety Intro Fellow and later an XLab member.

### Project Intern — C-DAC (Aug–Oct 2024)

Proof-of-concept deep-learning password tool using a GAN and a Transformer, with a PyQt interface and GPU support.

### Blaise Pascal Quantum Challenge finalist — top 15 of 137 projects (Jan–Mar 2025)

Modelled nitrogenase on a neutral-atom quantum computer, aimed at nitrogen fixation in ambient conditions, with a team across chemistry, quantum computing and computer science.

### BEng Computer Science, First Class — The University of Edinburgh (2022–2026)

Learned about various fields of CS, like Computer Architecture, Distributed Systems, Cloud Programming. Took some graduate courses.

### Exchange Student — University of Chicago (2024–2025)

Took graduate courses, including Fundamentals of Deep Learning at TTIC, contributed to XLab, and won in the course HFT competition.

## 05 — Contact

- Email: hello@arshkumar.com
- LinkedIn: https://www.linkedin.com/in/kumararsh/
- GitHub: https://github.com/thearshkumar