I'm kanishk, an undergraduate student in computer science.
I love working on LLMs and machine learning, making machines think using mathematics excites me a lot.
Currently exploring reinforcement learning, NLP, and computer vision. I like exploring rabbit holes about science, tech, and history.
Investigating whether a single transformer block reused recurrently can substitute for N independent blocks.
An API-first backend that gives LLMs structured context over databases.
Using language models as pure policies across classic Gym environments.
A lightweight parameter-efficient fine-tuning engine for practical model adaptation.
A football VAR system combining computer vision with multimodal reasoning.
want to talk or build something together?
mkanishkkulkarni@gmail.comResearch systems, model tooling, and applied products across LLMs, RL, and developer infrastructure.
An investigation into whether a single transformer block reused recurrently across N iterations with shared weights can substitute for N independent blocks while retaining competitive language modeling performance. Experiments across TinyStories and FineWeb-Edu reveal a narrowing perplexity gap with more training and a meaningful negative result showing test-time loop scaling degrades rather than improves performance.
An API-first backend for giving LLMs and agents structured database context. It connects to multiple databases, normalizes metadata into a unified model, builds a context graph, and adds embedding-based retrieval plus MCP-accessible tooling.
Explores using an LLM as a pure policy in classic Gym environments: CartPole, FrozenLake, and LunarLander. The model reasons over episode history in natural language with no gradient updates or fine-tuning, relying entirely on in-context reasoning to select actions.
A decoder-only Transformer built entirely from scratch in pure PyTorch. Implements modern architectural components — RMSNorm, Rotary Positional Embeddings (RoPE), SwiGLU feed-forward networks, and KV Caching, mirroring the design choices found in production-scale LLMs like LLaMA.
A lightweight parameter-efficient fine-tuning engine supporting LoRA, Adapter Tuning, Bitfit, and Prompt Tuning in a single unified package. Designed to make fine-tuning large language models cheap and accessible: installable via pip install tinypeft.
A Video Assistant Referee (VAR) system that analyzes football match footage to evaluate incidents and make foul decisions according to FIFA Law 12. Uses a two-stage pipeline combining computer vision for player and action detection with a multimodal LLM for objective, rule-aware decision-making.
A global disease surveillance dashboard that aggregates real-time outbreak data from multiple public health APIs and WHO feeds. Visualizes emerging threats on an interactive map, providing early-warning signals for pandemic preparedness and public health monitoring.
A multilingual AI assistant built specifically for Indian farmers, combining image recognition for crop disease diagnosis with real-time market price feeds from data.gov.in. Supports regional languages and covers crop advice, pest management, government schemes, and weather-based recommendations.
A reward model fine-tuned on Anthropic's RLHF preference dataset, designed to score model completions for use in RLHF training pipelines. Evaluated on the RewardBench benchmark to assess alignment quality across categories like safety, reasoning, and instruction-following.
Clean from-scratch implementations of core reinforcement learning algorithms: REINFORCE, Actor-Critic, and PPO — built in PyTorch with Gymnasium environments.
A collection of transformer architecture deep dives implemented from paper to code: KV Caching, Flash Attention, Rotary Positional Embeddings, and GPT. Each implementation is paired with a detailed blog post explaining the math and intuition behind the technique.
Llama 3.2 fine-tuned on curated physics question-answer datasets using LoRA and PEFT via Unsloth for efficient training. The resulting model demonstrates improved reasoning on physics problems spanning mechanics, electromagnetism, and thermodynamics compared to the base model.
A CLI security tool that scans web applications for exposed API keys, secret tokens, and sensitive endpoints in JavaScript bundles, HTML source, and network responses. Helps developers catch credential leaks before they reach production or get picked up by automated scanners.
A Model Context Protocol server that gives Claude and other MCP-compatible hosts the ability to read and write notes directly inside an Obsidian vault. Enables AI-assisted note-taking workflows, knowledge graph traversal, and automated linking between related notes.
Notes, explainers, and long-form writing on transformers, reinforcement learning, systems, and philosophy.
A high-level introduction to the core ideas behind agents, rewards, policies, and learning through interaction.
A practical breakdown of sparse expert routing, capacity management, and why MoE systems scale differently.
An accessible walkthrough of parameter-efficient fine-tuning and the tradeoffs behind adapting large models cheaply.
An implementation-oriented explanation of rotary positional embeddings with intuition for how they preserve sequence structure.
A comparative guide to how major transformer families differ in objective, architecture, and downstream use.
A guided tour through attention, feed-forward blocks, residual pathways, and the design logic of transformer models.
A conceptual bridge from formal computation to functional programming ideas that still shape modern languages.
A reflection on capability progress, open bottlenecks, and what it would actually mean to claim general intelligence.
A personal essay pushing against existentialist framing and testing its assumptions against lived experience.