# portfolio

Logs Discord messages, edits, deletes, reactions, and voice events to local SQLite, browsable via a real-time dashboard with full-text search.

Open-source Electron ebook reader for EPUBs and webnovels, with reading streaks, TTS, and Excalidraw annotations.

Ad-free YouTube Music player for your terminal, with a TUI and a CLI that stay in sync — so agents can control playback while the TUI runs in the background.

Fine-tune of TinyLlama-1.1B-Chat trained on chat logs to cultivate a friendly, witty Discord-style personality. Standalone merged model; superseded by v1.5.

v1.5 adds a second alignment stage using RLAIF with Gemini 1.5 Flash corrections — more knowledgeable, less prone to repetition, with a sharper persona. Merged, standalone model.

5M-parameter character-level GPT trained on TinyStories entirely on a Raspberry Pi 5 (CPU-only, ~9 days) — proof that capable LMs can be trained on consumer hardware.

Finetune of SmolLM3-3B-Base for casual, human-like, internet-native chat. Merged Float16 — QLoRA-trained, so don't re-quantize; best with an empty system prompt.

LoRA adapter for SmolLM3-Chat-v1 — load SmolLM3-3B-Base in 4-bit and attach the adapter for low-VRAM inference without the double-quantization noise.

Multi-label image classifier for tagging anime-style illustrations with a 10,000-tag Danbooru vocabulary. ConvNeXt V2 Tiny + GeM pooling at 448px.

Synthetic dataset (~20k samples) for training LLMs in chain-of-thought reasoning and tool calling. Chat format (system, user, model, tool), CC-BY-4.0.

Wuthering Waves English voice-line dataset: 37,805 samples, 913 unique speakers, 106 categories, ~24GB WAV with plain-text transcripts.