Google Reverse Proxy Permalink
Published:
OpenAI-compatible reverse proxy for Google’s AI products (Gemini CLI, Google AI Studio). Multi-account pool with quota balancing, failover, and a single bearer-token endpoint.
Projects
LumiBooks: Hierarchical Chat Memory Permalink
Published:
Two-tier in-place summarization for long chat histories. Older messages compress into ‘chapters’, chapters bind into ‘arcs’, summaries splice back where the originals lived so prompt order is preserved and context stays bounded. Stored as editable world-book entries with preview-before-save.
LumiAgent: In-Chat Tool-Calling Agent Permalink
Published:
A tool-calling LLM agent embedded directly in a frontend chat application, with full read/write access to characters, chats, prompts, personas, presets, and the assembled prompt. Every mutation flows through a diff layer with dependency-aware revert.
StructMemoryR1 — Tree-Structured Memory with RL Permalink
Published:
Three cooperating GRPO agents that build, query, and reason over a tree-structured memory of long multi-session dialogue. +9.0 F1 multi-hop and +13.8 F1 temporal vs. the strongest non-oracle baseline on LoCoMo.
LumiRealm: RisuAI Compatibility Runtime Permalink
Published:
An extension that ports the entire RisuAI character-card runtime (Lua triggers, ~179 CBS macros, lorebook decorators, fixed-position widgets, modules, in-app character browser) onto Lumiverse. Translator pipeline at import time, dispatch + sandboxed Lua at generation time, shadow-DOM-isolated rendering at paint time.
Lumiverse Permalink
Published:
Self-hostable AI chat platform with a sandboxed in-worker extension runtime. Bun + Hono + bun:sqlite (WAL, zero-ORM), Web Crypto for secrets at rest, native WebSocket event bus, and a typed RPC surface that lets third-party extensions reach chats, characters, generation, and macros under per-user scope.
LibreChat-Image Permalink
Published:
Personal fork of LibreChat adding standalone image-generation capabilities (NovelAI-style) and extra endpoint support on top of the upstream multi-provider chat UI.
Discord LLM Chatbot Permalink
Published:
Discord bot with multi-agent dialogue, vision, and per-channel chat memory. Wraps any OpenAI-compatible API so it can sit in front of self-hosted local LLMs.
Multi-Directional Ablation of Refusal in LLMs Permalink
Published:
Tests whether LLM refusal is mediated by multiple directions rather than a single one. Two-direction ablation outperforms the one-direction baseline — but not via the top-two ranked vectors.
ArtificialPersona Permalink
Published:
Real-time anonymous voice and visual communication built on speech-to-text, text-to-speech, and deepfake pipelines. Queen’s University team project.
DCP Tutorial #6 — Distributed Double Pendulums Permalink
Published:
Tutorial showing how to compute thousands of chaotic double-pendulum simulations in parallel on the Distributive Compute Protocol. Each worker computes one pendulum’s trajectory from a unique initial condition.
Distributed Newton Fractal Permalink
Published:
Computes Newton–Raphson fractals on Distributive’s web5 compute framework. Worker code is Python compiled to WebAssembly via Bifrost, so the same algorithm runs inside any browser-based DCP worker.