Work in progress · October 13, 2026

Hatsunama’s Free App Projects: Small-Model AI Research in Practice

Hatsunama, the online identity of Aaron Saunders, is a self-taught AI learner and developer building free, local-first app projects. Hatsunama’s independent AI research uses practical captioning and translation work to investigate small-model training, evaluation and the software around models. Current projects include BiliPop, Caption Studio, Caption Action, Local-only Finder and Sign to Speak.

BiliPop: local translation models and usable output

Hatsunama is moving BiliPop beyond its narrow English–Chinese OPUS setup toward the Caption Studio Hy-MT2-1.8B Q8 candidate, adapted from Tencent’s Hy-MT2-1.8B. Current work includes broader source- and target-language selectors, input/output layout cleanup and copying hashtags without adding extra hash marks.

Model downloading and behavior after stopping work or reopening the app are still being tested. The integration is unreleased; the public candidate evidence note separates mechanical checks from translation meaning, which still needs independent review.

Local-only Finder: practical local text search

Hatsunama’s Local-only Finder is an Android text-search project intended to be free. It searches selected UTF-8 notes and transcripts and shows exact source passages, using keyword search rather than AI-generated answers.

Recent work focuses on interface and keyboard repairs for typing searches and reading results. The repaired paths have been checked on two physical devices, with wider coverage pending.

Caption Action: testing live captions and audio capture

Caption Action is Hatsunama’s live-caption and translation app project. Current work investigates capturing the intended audio without audible speaker playback. A muted-audio diagnostic setup has been verified in a cloud environment; the app’s actual capture behavior there still needs to be demonstrated.

Sign to Speak: exploring on-device sign translation

Sign to Speak explores live-camera ASL/sign-language input to spoken English locally, with no recording or cloud upload as the intended design. This is research and prototype work: hand tracking alone does not establish sign translation, and reliable recognition has not been demonstrated.

Human First AI through free app development

Hatsunama’s Human First AI direction connects this work to useful software that preserves people’s meaning and leaves room for human review. The apps are intended to be free, with local-first behavior developed and checked feature by feature.

The research questions concern what training data small models need, how to evaluate translation meaning and how the surrounding app handles device limits and interruptions. Caption Studio’s public source, the translation model card and a proposed bilingual review method provide work readers can inspect.

The lessons feed into xMilo, Hatsunama’s larger mobile AI harness project.

Prepared with AI-agent assistance from current project updates and linked public artifacts. This is a development update, not a release announcement.