danyshs

"Golden Rule of Clear Prompting": Show your prompt to a friend, and have them follow the instructions. If they're confused, AI is confused.

—— Clear Prompting ——

Case 1 - Haiku
INPUT:
User : Write a Haiku

danyshs

NOblivion: Local LLM-Based File System Orchestrator

A privacy-first, "Sandbox-to-Disk" file organization and rule-enforcement engine.

[Current Status: Early Development / Prototyping]

Project Summary

NOblivion is a local-first utility that uses LLMs to bridge the gap between "fuzzy" organizational desires and rigid file system structures. It allows users to define, test, and enforce complex directory rules through natural language. Its standout feature is a Virtual Filesystem Sandbox, where users can visualize and iterate on a proposed organization before any actual move or copy operations are executed on the host machine.

Why This Project Matters

File organization is often a choice between manual tedium or "black-box" automation. Most users "lose" files to digital oblivion because their organizational needs change over time, but the effort to restructure thousands of files is prohibitive.

danyshs

RWKV-7: A Specialized Approach to Symbolic Piano Continuation

Official Submission for MIREX 2025 Symbolic Music Generation Challenge

Read the Paper (arXiv) | Read the Paper (ResearchGate) | GitHub Repository

Project Summary

This research explores the efficiency of the RWKV-7 "Goose" architecture in the context of symbolic music. The task—taking a 4-bar piano prompt and generating a coherent 12-bar continuation—is often approached with massive Transformers. This project proves that a 20M parameter RWKV-7 model, trained on high-fidelity data with a lean tokenization strategy, can match or outperform models with hundreds of millions of parameters.

The Engineering Thesis: "The Hammer vs. The Mallet"

Why use a massive general-purpose Transformer for a task a specialized decoder can handle more effectivel

danyshs

DiscWhisper: Local-First Multi-Speaker Live Captioning

A high-performance accessibility tool for real-time, color-coded speech-to-text.

[Current Status: Proof of Concept / Active Development]

Project Summary

DiscWhisper is a local LLM-powered captioning system that processes live multi-speaker audio (e.g., Discord, meetings, or media) and generates visually distinct, color-coded transcriptions. Unlike standard closed captioning, DiscWhisper utilizes Speaker Diarization and Active Learning to identify who is speaking, assigning each voice a persistent visual identity.

The Idea

Overstimulated? Differently abled regarding audio-processing? Do you get lost in the noise the moment you have more than 3-4 people talking?

danyshs

CareGiver: AI-Driven Assisted Living Management

Winner of the DeveloperWeek 2025 Hackathon (1st Place: Oracle Challenge & PliOps Challenge)

View Devpost Project | View Source Code

Project Summary

CareGiver is a dual-interface application designed to modernize assisted living facilities. It bridges the gap between high-friction administrative data entry and resident cognitive health. Built using Oracle APEX and a custom Python/Redis backend, the system features:

1. An AI "Super-Secretary" for Admins: Converts natural language into database actions to eliminate data silos caused by staff turnover.
2. A Resident-Facing Cognitive Assistant: A specialized chatbot (DeepSeek-R1-Distill-Qwen 1.5B) trained on medical communication data to provide health information and "Historical Persona" cognitive games.

danyshs

Local-First Personal Wiki RAG

A privacy-centric, local LLM Retrieval-Augmented Generation (RAG) system designed to semantically index and query personal knowledge bases (Markdown and Neorg files).

Project Summary

This project acts as an interactive layer over a personal, plain-text wiki. It ingests daily notes and documentation, stores them as vector embeddings, and exposes them through a chat interface. Unlike standard LLM wrappers, this system is heavily constrained to prevent hallucinations: it is prompted to strictly cite its sources down to the exact file name, line number, and exact quote. Everything runs locally, ensuring 100% data privacy.

Why This Project Matters

Note-taking systems inevitably become massive and unwieldy. While traditional search (grep) is precise, it lacks semantic understanding. Cloud-based AI tools solve the semantic problem but introduce unacceptable privacy risks for personal data and a high tendency to hallucinate.