Samit Mohan
I'm an AI Scientist at Beltech AI in Bangalore, building computer-vision systems for city traffic: live CCTV, ≤10ms inference, 0.9+ F1 in production. In my free time I rebuild ML from scratch and write engineering blogs.
From Scratch
View allThe math behind deep learning - gradients and the chain rule, from scratch
Hello world of neural networks - a digit classifier in pure NumPy
NumPy + PyTorch from scratch - tensors and autograd, then the same net in PyTorch
Building PyTorch from scratch - a working framework in ~1300 lines
Attention residuals - why attention and residual streams work
The annotated microGPT - a GPT built and trained end to end
What happens when you press 'submit' on ChatGPT - the full pipeline: data, alignment, inference
Inference engineering - the KV cache, batching, what makes tokens fast
Projects
View all40M GPT trained from scratch on math. Custom BPE, GQA + RoPE + SwiGLU, sparse MoE, SFT + GRPO reasoning RL. Evaluated on GSM8K.
PyTorch in ~1300 lines. Reverse-mode autograd, conv2d, optimizers, LR schedulers. 96%+ on MNIST.
Built to figure out what's actually inside an agent loop. Turns out it's 250 lines: tool calls, shell exec, streaming.
YOLO + ResNet on match footage. Court detection, ball tracking, shot classification. Tested on Wimbledon clips.
Experience
Beltech AI ・AI Scientist
City-traffic computer vision on live CCTV: ≤10ms inference, 0.9+ F1 in production, 100+ cameras per server.
Genpact ・Data Science Intern
Built an NLP resume screener for data-team hiring and a churn-prediction model (POC), plus dashboards and ETL. Alteryx certified.
MakeMyTrip ・Machine Learning Intern
Personalization API with multi-armed bandits and an RLHF reward model for itinerary ranking.
Education
NHCE, VTU ・B.Tech Information Technology
CGPA 9.4 / 10 · Founder, Coding Club NHCE.