P3 RELOAD // SYSTEM
SYSTEM RELZZY'S ARCHIVE // PROJECTS
1-7 QUICK SELECT  |  CLICK SELECT PROJECT  |  ESC MAIN MENU

PROJECTS

07 ITEMS
B RETURN TO HOME

Code Intelligence & RAG

One of my main projects is a local code intelligence and RAG system built around a fine-tuned Qwen3.5-2B language model. The goal is to create a lightweight AI system that can understand a large codebase, retrieve the right information, and interact with the local Linux environment when needed.

The model itself is fine-tuned specifically for Linux-focused tool use and routing, rather than being used as a general-purpose chatbot. It learns to determine what the user is asking for and return a structured tool call that can be handled by the rest of the system.

How the system works

  • Fine-Tuned LLM Router — A 2B-parameter Qwen3.5 model fine-tuned using QLoRA and Unsloth. It classifies the user's intent and produces structured JSON tool calls. The training data is focused on Linux commands, system operations, codebase queries, and tool usage.
  • Codebase Retrieval — When a question requires understanding the project itself, the request is sent to the RAG pipeline rather than directly to the terminal.
  • Hybrid Search — The RAG system combines dense vector search using ChromaDB with BM25 keyword search. This allows it to find both semantically related code and exact matches such as function names, variables, filenames, and technical terms.
  • Reciprocal Rank Fusion — Results from the different search methods are combined using RRF (Reciprocal Rank Fusion), producing a stronger combined ranking instead of relying on a single search method.
  • Cross-Encoder Re-ranking — The most relevant results are then passed through a cross-encoder reranker to improve the final selection of context given to the language model.
  • Linux Tool Integration — System-level questions can be routed to controlled Linux tools for things such as file discovery, logs, processes, memory usage and system inspection.
  • Final Response — The retrieved code or system information is passed to an LLM to produce a grounded response based on the available evidence.

The system has been tested on a codebase containing 350+ files, successfully indexing and retrieving relevant information from across the project. The initial indexing is a one-time process, after which queries can use the existing search indexes instead of processing the entire codebase again.

The trained model can also be deployed locally using 4-bit GGUF quantization, making the system practical on consumer hardware while keeping the data and inference local.

⟠ View Source Code 🤗 View Fine-Tuned Model

Full-Stack Web Development

I've also built and deployed several full-stack web applications rather than stopping at the frontend. These projects have involved designing interactive interfaces, building backend functionality, handling user data, implementing authentication, and deploying the complete application.

  • Custom backend functionality for forms and user interactions.
  • Firebase for storing submitted data and managing application data.
  • Google and Discord authentication with protected user areas and working login systems.
  • User dashboards with personalized content and account-based functionality.
  • Deployment using services such as Render, allowing the applications to run as actual hosted web services rather than just local demos.
⟠ GitHub Profile

AI-Driven Narrative RPG

One of my newer projects is a web-based RPG that combines traditional game systems with AI-generated content. Instead of allowing AI to completely control the story, the game uses a fixed canonical main storyline as its foundation while allowing AI to generate flexible dialogue and events around it.

  • AI-generated dialogue and procedural events that respond to the player's situation.
  • A persistent character, inventory and statistics system that keeps track of the player's state.
  • Player decisions and game state can affect future events and potential endings.
  • The AI operates within the game's existing rules instead of being allowed to freely rewrite the underlying game state.
  • The project is currently in beta and runs through a web interface.

Visual Novel / RPG Engine

I also built a reusable visual novel-style game engine rather than creating only a single fixed story. The engine separates game logic from narrative data so that new stories and content can be added without rewriting the entire system.

  • Save and load system for persistent player progress.
  • Character appearance and wardrobe systems that can influence the game and narrative.
  • Persistent stats, money, inventory and time systems.
  • Branching story logic where choices and player state can affect later events.
  • Designed to work across both PC and mobile layouts.

The Last Divine — Text RPG

One of my earlier projects was The Last Divine, a text-based RPG built from scratch with Vanilla JavaScript and JSON. I designed it around a modular game system rather than a simple collection of "choose an option" pages.

  • Persistent Save/Load system using localStorage.
  • Inventory and item management, including usable items and consumables.
  • Health, energy and character statistics with conditional gameplay logic.
  • A reactive appearance system where equipped items can change how parts of the story are displayed.
  • Narrative and game data separated from the core game logic for easier expansion.

Game Development & Storytelling

Outside of programming, I'm also interested in game development and storytelling. I write story scripts and design characters, worlds, gameplay systems and narrative structures for games I want to build in the future. My long-term goal is to create a 3D indie game that combines interactive gameplay with a strong story.

Web3 & Community

I'm also active in decentralized and Web3 gaming communities, where I've participated as a community member, creator, volunteer and player. I've contributed to communities around decentralized games and projects and have been involved in helping promote and represent projects online.

One particularly memorable part of that experience was being an early member and creator within a decentralized gaming community and receiving sponsored flights to Mumbai on three separate occasions to attend creator meetups and represent/promote the game.