Final-year CS student at SIT (MAKAUT) . Research assistant, University of Manchester . Builder of things that probably shouldn't work but do.
I got into programming at 14 through security research -- no proper setup, just constrained hardware and too much curiosity. That spirit of figuring things out from first principles is still what drives most of what I build. Most of my work has and continues to come from me being frustrated at the work done by other people or at the fact that there was a lack of some specific thing and no one bothered to work on it or fix it.
Right now I'm somewhere at the intersection of AI research and systems thinking, trying to answer a question I find genuinely interesting: what does it actually take for an AI to understand context the way a person does?
AI-Player -- A Minecraft mod that's less of a "bot" and more of an experiment in agentic AI. It has a dual-brain architecture: a reactive reflex system for fast decisions, and an LLM-backed reasoning layer for higher-order planning. Custom NLP pipeline (DistilBERT + CART + LIDSNet), risk-augmented Q-learning, RAG systems, multi-LLM provider support. Minecraft is a surprisingly good testbed for AI research -- open-ended, dynamic, and forgiving enough to iterate fast.
Lexis -- A classical text compression system with zero training data, reaching ~2.75 bits per byte on real-world web documents. Outperforms gzip and zstd. Built on eight linguistically-structured stages including mixed-radix phonetic decomposition and factoriadic delta encoding. Paper in progress for arXiv.
Mycelium (A new AI architecture) -- My final year project. A multi-model AI architecture designed around how the brain might actually handle knowledge -- continuous learning without catastrophic forgetting. Less about scaling up, more about getting the structure right, currently in the works. Aims to solve a lot of the problems current AI systems with LLMs face.
Adaptive Memory Integrity & Telemetry Framework (AMITF) -- AMITF is a research framework aimed at defeating cheating in Counter-Strike 2 (and similar multiplayer games) through memory fragmentation, encryption epoch rotation, and behavioral telemetry -- without relying solely on static signatures or kernel-level enforcement. I decided to start working on this because I was done facing cheaters in Counter-Strike 2 and also waiting for Valve's "wait-and-see" approach, too slow for me.
CSE Engineering Notes -- A free, handcrafted note repository for CS students. Made it because the good stuff was always scattered across a dozen sources. Now it's on Obsidian Community Vaults.
Small libraries I published 4 years ago that apparently kept finding users.
- requests_rs -- Python
requests-style HTTP ergonomics for Rust · 11k+ downloads - minecraft_launcher_core · 4.5k downloads
- minecraft_downloader_core · 10k+ downloads
Research assistant (remote) with the University of Manchester -- working on AI systems and agentic architectures. I've spent time on NLP pipelines, retrieval systems, reinforcement learning, and compression theory. Most of my instincts around systems design come from years of building on hardware that couldn't afford inefficiency.
Java · Python · Rust · C · JavaScript / React / Tailwind · HTML / CSS
ML/AI: NLP · RL · Deep Learning · RAG · Model training · Knowledge systems
Other: Game modding · API integrations · Distributed design · Linux (Nobara)
I'm interested in research roles, collaborations, and conversations about AI that go deeper than the surface level.
