I build neural networks from scratch and write down exactly how they work.
Research prototypes to production pipelines. Most of it runs on hardware you probably already own.
|
Text, image, audio and video in one shared 768-dim embedding space, trained end to end on a single RTX 4060 (8GB). An independent from-scratch reproduction of the GELATO architecture from Jina AI.
|
Self-Adapting Language Models in a single readable file. The model writes its own fine-tuning data, tests whether it actually helped, and reinforces the habit. ReST-EM on 8GB, degrades gracefully to CPU.
|
|
Gives Claude Desktop eyes through your webcam. Local vision via Ollama, temporal memory of the last five minutes, no cloud vision APIs. An MCP server that installs in three steps.
|
Engineering governance at the one place every change passes through: the commit. A managed global Git hook that stops credentials entering history, across every repo on a team.
Python port. Original design and detection engine by Gaurang Joshi, Shrey Tandel and Dwij Acharya. |
|
The invisible fingerprinting technique Netflix and music labels use to catch pirates, implemented cleanly in Python. Embed, detect, and everything that breaks it.
|
An end-to-end edge detection pipeline: train, optimise, deploy to a Raspberry Pi. Native Keras and KerasCV, no Ultralytics licence required.
|
| MODELLING | |
| AGENTS | |
| PRODUCT |
- **[I Built STAIR From Scratch: A Retriever That Returns Section Names, Not Chunk IDs](https://medium.com/@mithilmaske/i-built-stair-from-scratch-a-retriever-that-returns-section-names-not-chunk-ids-22882a7b4cd8?source=rss-8d19d7c20101------2)**
September 20, 2026
- I Built a Multimodal Embedding Model From Scratch on an RTX 4060 (Text, Image, Audio, and Video…
August 22, 2026 - How I implemented Self-Adapting Language Models with LoRA and QLoRA on an RTX 4060
July 25, 2026 - How to Build a Vision Language Model from Scratch Using Q-Former, Contrastive Learning, and LoRA
June 27, 2026
Open to research collaborations and applied AI work — reach me on LinkedIn.


