Last Updated: August, 2025
I am Sankalp. I am into AI engineering and LLM research (mostly applied for now). I have experience in backend engineering, crafting products, context engineering, LLM evals, agentic code generation, voice agents, model behaviour and a bunch of other areas. It's already many buzzwords but believe it or not, AI in itself is very broad now even just on application side of things.
I am open to Consulting work or just chatting on interesting stuff if time allows. You can DM me on X. Leads are appreciated.
I stay up to date with latest AI progress and love hanging around with my mutuals on X tracking the same. It's fun plus helps me in my job. If you wanna stay up to date, I have curated a list on X from where you could follow some people and see who interacts with them etc.
I love to play around with models and writing which contribute to my intuition and tacit knowledge around these. My other interests span - distributed systems broadly, post-training broadly, role playing, inference optimisation, search and retrieval, RL environment, and recently have started dabbling more into pre-training research.
I love using AI devtools like Cursor, Claude Code and have used/reviewed several products either for personal usage or as feedback to companies/friends. It's a pity though that I haven't built much in this area in 2025 - unlike last year.
I like to write, shitpost, occasionally watch anime, play table tennis, hanging around with friends in Bangalore's cool weather. Would love to play more video games (Elden Ring is my #1). My favorite band is Men I Trust.
Popular technical posts
My Claude Code Experience after 2 weeks of adventures - This is a guide on Claude Code. This also reached Hackernews Front Page
Evolution of AI Assisted Coding and Developer Interaction Patterns - Get a glimpse into how we went from generic autocomplete to current coding agents and tips to vibe code better. Windsurf CEO Varun Mohan loved this blog!
Shape Rotation 101: Intro to Einsum and Jax transformers - This post reached Hackernews frontpage!
An Attempt To Build Cursor's @codebase feature
Learning ML in 2024 - Check out the recommended section
Popular non-technical posts
Projects [Outdated]
This section showcases side projects I built for fun and learning last year.
CodeQA
CodeQA is an attempt to build Cursor's @codebase feature. It allows users to search the codebase and get relevant files, code snippets by asking questions in English (natural language). It can support multiple languages.
It utilizes tree-sitter to parse AST and construct a codebase index before generating embeddings. We do a top-K RAG (Retrieval Augmented Generation) using these embeddings along with more post-retrieval techniques. I wrote a couple of blogposts (that are listed in the above section) for CodeQA.
Note: CodeQA works 2.5x faster than what's shown in the video now. I updated it later. Check this branch
AutoLatex
Autolatex (400+ users so far)
AutoLatex is a browser extension that simplifies LaTeX equation generation for researchers and students. It uses LLMs to convert natural language and images into markdown and LaTeX, with instant rendering so you can edit on the spot. The tool offers customizable settings, keyboard shortcuts, and privacy control through user-managed API keys. By streamlining the process of incorporating complex equations into documents, AutoLatex enhances productivity in academic and research workflows, especially in deep learning fields.
VimSonnet
VimSonnet was a small project where I tried to control my browser using Claude 3.5 Sonnet as agents. Github
The approach in the demo uses Vimium (a Chrome extension that generates hint tags for all UI elements in a browser) along with Pyautogui and Claude 3.5 Sonnet as the LLM.
I had tried a screen coodinate based approach but Claude's Vision cannot point out coordinates by itself so it was required to use hint tags.
all i wanted to do was post a tweet by claude and that has been done. the screen coordinate approach doesn't work unless your vision model supports coordinates (some can do, some getting cooked). this approach uses vimium hint tags + pyautogui + 3.5 sonnet function calls https://t.co/w32UBKBM60 pic.twitter.com/8tCYDMLEhT
— sankalp (@dejavucoder) June 30, 2024
Twitter Circle
A tool to visualize your Twitter network and direct messaging history.
- Make a Twitter Circle visualization for up to 200 users.
- Check leaderboard based on combined weights of all your mentions of other users and all direct messages.
- Check DM stats message count per recipient, messages sent/received per user, total messages, last message with them
- DM bar graph where you can see messages/month for 5 years data.
This project has been used for 200+ people by now and recieved 100+ stars.
https://t.co/gVQAaUelSC pic.twitter.com/EKRQYPkjDC
— sankalp (@dejavucoder) June 14, 2024
SemanTweet Search
SemanTweet Search allows you to search over all your tweets from the Twitter archive using semantic similarity. A demo is available here.
It preprocesses your tweets, generates embeddings using OpenAI's small/large embedding model, stores the data and embeddings in LanceDB vector db, and provides a web interface to search and view the results.
You can do semantic search post pre-filtering by time, likes, retweets, media only or link only tweets too.
Pre-filtering by sql operations helps not only filter but also reduce the vector search space thus speeding up the search.
You can additionally use/edit projector.py
 and tensorflow projector to get a visualization of your tweets using t-sne algorithm as shown here
Demo Thread Visualization Demo
Handpicked by Haiku
A LLM based app that generates personalized recommendations and visualizations based on user input, leveraging Claude Haiku's knowledge base.
CaptionBot
This is my Pytorch implementation of the paper "Show, Attend and Tell".
It generates descriptive image captions for your images. Architecture involves Seq2Seq-based image captioning with attention mechanism, utilizing ResNet-50 as encoder to extract meaningful image features and an LSTM-based decoder with soft-attention and Beam Search. It achieved BLEU-1 score of 59.2 and BLEU-4 score of 19.56, closely matching the original benchmarks