Testing ideas before they become products.

Exploring

A Chrome extension prototype for adjusting volume and playback speed on supported audio and video in the current tab. It is being tested and is not available in a store yet.

Current status

Version 1.1.0, an internal verification candidate. Store publication has not started yet. Verified on Windows + Chrome; other operating systems and sites are unverified.

Exploring

Which local runtime works best for Qwen3.8-27B on one RTX 5090? We compare inference speed and coding-task behaviour, then publish the settings and measurement conditions.

Why it matters

The data points to different strengths rather than one universal winner. Q5/MTP3 decodes roughly 1.5–2× faster in bounded runs, while SGLang provides stable concurrent serving at 80K+ context. Both runtimes passed the correctness gates: basic 6/6, JSON schema 20/20, tool calls 40/40, NIAH at 32K/80K/113K+ 5/5 each, with zero CUDA/OOM/output-corruption events. Raw decode TPS therefore does not represent agent throughput on its own.

Current status

The final production speed ranking awaits comparable end-to-end wall-time measurements across multiple seeds, a limitation documented in the repo. The September 2026 NInfer comparison is complete and a Codex failure is documented; Q5 remains the default single-agent runtime while the wall-time comparison is pending.

Exploring

Can a writing tool keep characters, past events, and world rules consistent over many episodes? We are developing a system that checks new text against the story's recorded state before carrying it into the next episode.

Why it matters

A language model has no persistent memory of the world it is writing. After a few scenes, personality drifts, earlier resolutions get forgotten, and spatial or temporal continuity breaks. Keeping a large, consistent world state inside a finite context window — while still producing natural prose — is the hard part.

Current status

This is active research, not a released product. We'll share more of the approach once it's stable enough to describe accurately.

Exploring

Can a small model running locally help an engineer interpret messaging-system logs without recommending unverified or destructive changes? We are testing training and evaluation methods for that task.

Why it matters

Enterprise message brokers involve complex failure cascades — thread pool exhaustion, broker store limits, corrupted persistence indices, and cascading timeouts. Small language models tend to hallucinate plausible-sounding configurations or suggest dangerous administrative operations without isolating the root cause.

Current status

This project is strictly private exploratory research, not a commercial product or automated operations agent. All experiments are conducted in isolated offline environments without live production access.