Harshal Singh
Last updated July 2026
Abstract
I am an independent researcher and engineer.
My broader interests are computer science, mathematics, physics, and hardware. I learned almost everything I know from the internet.
I believe in techno-optimism and effective accelerationism. The plan is to build profitable companies that accelerate human progress.
I believe the world is a very malleable place, and that I can bend it according to my will.
1. Operating principles
- Move quickly; make plans for months and complete them in weeks.
- No risk, no gain.
- Do something that normal people cannot even dream of.
- The opposite of success is mediocrity, not failure.
- This time will pass; you will get old, you will die, so it does not matter.
- No one is rich in time.
- Competition is for losers.
- The laws of physics are the only limit.
- Model the world as 500 people rather than 8 billion.
2. Current Work: Sovereign AI for Inference and Model Training
Most of my time goes into:
- Training small language models and beating them on benchmarks.
- Taking open source models, often from China, and pushing them to run under extremely constrained compute budgets.
- Fine-tuning open models to outperform frontier models on specific tasks.
AlphaSpectra: Detecting Plant Disease Before Symptoms Appear
Developing a foundation model trained on hyperspectral leaf imagery to detect plant diseases before visible symptoms emerge. The model identifies spectral changes that occur before any signs can be seen by the human eye.
The first public release is BridgeCheck, a physics-grounded VNIR-to-SWIR candidate generation system with fail-closed auditing on paired measurements. It ensures that a predicted value is never silently presented as a real measurement, preserving scientific reliability and traceability.
DM for any type of model training/inference work. X or email.
3. Fun facts
- 21. College dropout, in tech for the last 5 years. Used to run gladium.ai and scaled it to $10K.
- Built a product for landlords and tenants that reached over 100,000 users.
- Scraped the top subreddits, built a deep-research system to analyse them, and published results that crossed 200,000 views. A creator with 700k followers reached out.
- Trained a Hinglish TTS model on synthetic data that beat all the open-source models.
- Off-screen: I run, and I box (beginner).
4. Writing
Essays at the intersection of work, love, philosophy, and future technology. View all essays →
5. Past experiments
- 5.1 Humanless
- A plug-and-play AI tool that transforms existing CCTV footage into actionable business insights. Connect existing cameras, analyze in real time, get reports. Aimed at retail and hospitality.
- 5.2 IFarc
- AI-powered compliance automation for Registered Investment Advisors: automated recordkeeping (Rule 204-2), code-of-ethics monitoring, client onboarding, and marketing review for SEC compliance. The teacher here was the regulated industry itself.
- 5.3 AI Video Creation Platform
- An experiment in automating high-quality YouTube video creation end to end. Prompt in, ready-to-publish content out. Spun off three open-source siblings: Open Podcast (text to podcast), Open Dubbing (video dubbing & audiobooks), and Connected Podcast (AI-simulated conversations with historical figures).
- 5.4 Historical Conversations
- Imagined conversations with Ramanujan, Steve Jobs, and Nikola Tesla. Listen on Spotify →
- 5.5 AI research
- Working on a novel model architecture. Read the thread →
6. Projects & code
| barunlm-35m | The world's best language model under 100M parameters: 41.01% macro accuracy on a decontaminated nine-task zero-shot suite at 35M parameters, beating LFM2.5-230M-Base with 6.55x fewer parameters. |
| syntts | SOTA Hinglish (89M) code-switching TTS running fully in the browser on WebGPU; trained on 100% synthetic data, distilled from 467M with RL. |
| alphaspectra-bridgecheck | Browser-first, physics-grounded VNIR-to-SWIR candidate generation with fail-closed paired-spectrum auditing for leaf spectroscopy. |
| tinyrouter | A ~10K-parameter LLM router that learns which open-source model should answer each question and in what role, trained by evolution (sep-CMA-ES). Beat every single model in the pool; picked up by researchers at OpenAI and Anthropic, and reposted by the CEO of Hugging Face. |
| numpygrad | GPT-2 trained from scratch in pure NumPy, verified against PyTorch, with an accompanying blog post deriving the math line by line. |
| twitter-memory | MV3 Chrome extension that captures your Twitter attention log (tweets, sessions, dwell time) into a local SQLite DB with daily markdown exports. |
| postforge | Self-improving LinkedIn content engine: multi-agent simulation, EMA weight calibration, 10-persona debate threads, cron automation. |
| skore | Production-grade chatbot for JEE/NEET students, trained on YouTube videos and Reddit threads. |
| mini_alphafold | A minimal implementation of AlphaFold for protein structure prediction. |
| ask_shr_krishna | AI-powered Bhagavad Gita Q&A system. |
| picrate | An appearance analyzer, an AI tool that scores facial images. |
| gpt2-accelerator-chip | Designing a GPT-2 accelerator chip from scratch in under four weeks. |
| blogify | Transform podcasts into engaging blog posts. |
| CalorieCam | Count calories from food images using computer vision. |
| webscribe | Chat with websites interactively. |
| DataDigitalizer | Bridge physical and digital data. |
| LLAMA_-from_scratch | Implement LLAMA from the ground up. |
| Open-Dubbing | Open-source video dubbing and audiobook creation tools. |
7. Other
Note
The two current limitations of LLMs, as I see them, are research taste and system design.
They can execute, but they cannot tell which problem is worth picking up, and they fall apart the moment a problem requires holding a whole system in mind at once.
My bet is that the underlying mathematics is essentially solved. The arbitrage left on the table is in taste and architecture: in choosing the right problem and composing the right system around it. That is where I am pointing my work.