My Projects

Here you'll find a collection of projects and experiments I've worked on, ranging from web development and open-source tools to personal learning explorations.

Research Work

• Object Hallucinations in Multimodal Large Language Models: A Survey [github]
  Multimodal Large Language Models (MLLMs) have shown remarkable advancements in tasks that integrate textual and visual data. However, they often generate outputs that are inconsistent with the provided visual content, raising concerns about their reliability. This survey aims to provide insights into the current research landscape regarding hallucinations in MLLMs, integrating existing knowledge and identifying future research directions.
• Relation Extraction using LLMs [github]
  A survey of relation extraction methods, focusing on their evolution, challenges, and future directions. It covers traditional methods, deep learning approaches, and the latest advancements in the field.

Projects

• Cairn [Source]
  Cairn is a Chrome extension for reading the web with something left over. Highlight a passage, write a note, or save a whole page as you browse. No folders, no filing, just one archive. Captures write to local storage instantly and sync afterwards, so a bad network never costs you a save. The archive is also exposed over an MCP server, so Claude, Cursor, or any MCP client can search what you've read mid-answer instead of that context being stuck inside one assistant.
• Nanostore [Source]
  Nanostore is a vector store built from scratch to find out where binary quantization's speedup actually comes from. It implements the whole path (chunking, embedding, storage, search) and then adds the optimization: 1-bit sketches filtered by Hamming distance, reranked with exact cosine. The 32x index compression showed up as promised. The speedup didn't. Across a sweep from 1k to 100k vectors the bit filter ran consistently slower than plain brute-force cosine in NumPy, which puts the real win in SIMD popcount rather than in the funnel itself.
• Veritas
  Veritas is an AI research engine that makes research faster, deeper, and more trustworthy. We decompose papers into their smallest indivisible claims, map the logical dependencies between them, and run each through symbolic verification and adversarial stress-testing. Every claim either holds or gets flagged with evidence, not opinion. Research deserves better infrastructure. We're building it.
• Mycelium [Demo] [Source]
  Mycelium is a persistent memory for web agents. Mycelium analyses every run of your AI Agents then figures out what it could do better next time, then it injects the hints into the prompt next time your agent runs. This slowly makes web agents faster and more efficient at accomplishing their tasks as they are less likely to get stuck on repetetive blockers.
• Heph[AI]stus [Demo] [Source]
  HephAIstus is an autonomous coding assistant powered by ReAct (Reasoning and Acting) methodology. Named after the Greek god of fire and forge, HephAIstus helps you build, modify, and manage code projects through natural language conversations.
• Implementing Deep learning Research Papers [Source]
  A collection of implementations of various deep learning research papers. The goal is for me learn and understand the concepts behind these papers by implementing them from scratch.
• VGrad [Source]
  An autograd engine that computes gradients of functions using reverse mode differentiation. It is designed to be simple and easy to use, making it suitable for educational purposes and small-scale projects.