DEV Community: Wal

serifu: translating manga in place on android

Wal — Thu, 05 Nov 2026 15:22:44 +0000

serifu is an android app that translates manga, manhua and manhwa in place. it screenshots the reader app you choose, finds the speech bubbles and draws english over them. no exporting pages, no pc pipeline.

how it works:

an accessibility service takes the screenshot when you tap a floating pill (or automatically when the page stops scrolling)
the page gets split, bubbles get clustered, ocr reads them, the text gets translated and lettered back over the bubbles
translation is on-device with ml kit, or an optional openai-compatible cloud endpoint if you want a stronger model
the reading loop is a reducer: accessibility callbacks in, next state plus effects out

the app is three gradle modules: :core (plain kotlin), :app (the phone), and :linen (the design system the screens are built on).

it bundles the full volume 1 of give my regards to black jack (209 pages, shared by the author under free secondary-use terms) so there's a sample to translate out of the box.

it's 0.6.7-beta, still early development, and ocr and lettering will keep changing between releases. source-available under polyform noncommercial, not osi open source.

code and apk: https://github.com/walsoup/Serifu

[Boost]

Wal — Thu, 22 Oct 2026 23:58:28 +0000

taha

Oct 22

SlothTidy: A Zero-Friction, 100% Local AI File Daemon That Organizes Your Downloads While You Rest 🦥

#hacktoberfest #opensource #ai #python

5 min read

paste-sloth: Ctrl+Space+V pastes your clipboard back tidied, by a 1-bit Gemma

Wal — Thu, 22 Oct 2026 23:50:00 +0000

Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady Submission 🦥

This is a submission for the Hacktoberfest Open-Source AI Challenge: Week 3, Slow and Steady.

What I Built

paste-sloth is a background process. you copy the way you already do, and when the text is a mess you hold ctrl, press space, press V, and let go. the clipboard goes to a model, comes back restructured, and gets pasted into whichever window has focus. csv becomes a markdown table, json gets pretty-printed, logs become one entry per line.

name,score,city
ada,10,lagos
grace,12,accra

the usual repair is a second app: paste into a chat, type "just the table", wait, select the reply, copy it, go back, paste. that whole loop is now one chord. the chord is awkward on purpose, so it never fires while you're typing, and the model is small enough that leaving it running feels lazy rather than expensive.

Demo

the gif is a real paste-sloth clean run against a stub server. the stub is rude on purpose: it answers with Here is the table: and a code fence, so you can watch the tool strip both, check the numbers, and print the table. paste-sloth clean restructures the clipboard once and prints the result - the easy way to see what the model does before the keyboard is involved.

the installer is one line. it finds python 3.11 or newer, makes a venv, and opens a menu:

# macOS or Linux
curl -fsSL https://raw.githubusercontent.com/walsoup/paste-sloth/main/install.sh | bash

# Windows
irm https://raw.githubusercontent.com/walsoup/paste-sloth/main/install.ps1 | iex

enter on the first row asks for a base url, a key, and a model name. down, then enter, is the local option: it installs ollama if it's missing and downloads gemma 4 e2b at 1-bit (IQ1_S, about 2.4 gb - google doesn't publish a 1-bit file, so it's an importance-matrix quant of the instruction-tuned weights), then writes the config and an autostart entry.

How I Built It

the model does the one job a very small model can be asked to do. the code holds the pen for everything else.

the request is a chat completion at temperature 0. the system prompt is short: reply with only the restructured text, keep every value, pick the structure that fits. before the reply touches the clipboard it gets checked:

code fences and one-line preambles ("here is...") are stripped
if the reply contains a number the input never had, it's asked once more with a sterner prompt, then dropped. invented numbers are the worst failure for pasted text: a figure that looks right and isn't. dropped numbers are allowed, because stripping timestamps off a log is legitimate
empty reply, error, or timeout (15s by default): nothing is pasted. your original text is still on the clipboard

the number check is the part i would not ship without:

def introduces_numbers(source: str, output: str) -> bool:
    """True when the output contains a number the input never had.

    Dropped numbers are allowed: stripping timestamps from a log is legitimate.
    Ordered-list markers the model adds (``1.``, ``2)``) are not data.
    """
    invented = number_tokens(output, ignore_list_markers=True) - number_tokens(
        source, ignore_list_markers=False
    )
    return bool(invented)

the chord was the fiddly part, because space is also a spacebar. ctrl is forwarded as it happens. space is held back, and if you press another key or release space before V, the held space gets replayed, so autocomplete and input-method switching still work. V while both are held is the paste, and both keys are swallowed so the app doesn't also get a raw paste. a second chord while a request is in flight is ignored. ctrl+space+esc releases the keyboard and quits.

on linux this is evdev with an exclusive grab plus a uinput virtual keyboard - wayland won't let an ordinary app watch every key or type into other windows, and a desktop shortcut can't express ctrl+space+V because space is not a modifier. the grab needs a one-time sudo setup, which paste-sloth setup prints. windows and macos use pynput.

Why Does Open Innovation Matter?

the clipboard is the most private stream on the machine: passwords, tokens, a paragraph you haven't sent. "reformat what i just copied" is the last feature i'd hand to a service i can't inspect. with ollama on localhost, nothing you copy leaves the machine.
open weights are what make this version possible. google released gemma, someone quantized it to 1 bit and published the file, ollama serves it on cpu with no account. every layer can be read, swapped, or run with the network off.
it costs nothing to run, so there's no meter on a chord you hit all day.
not locked in: base_url takes any openai-compatible endpoint, local or hosted. your clipboard, your pick.

what it can't do yet

windows and macos are written and unit-tested, but not run on real machines. macos will ask for accessibility and input monitoring, and ctrl+space may already switch input sources.
on my own wayland box the uinput paste still needs the sudo setup, so a live ctrl+v from the virtual keyboard is unconfirmed there. wl-copy and wl-paste do work from a detached process, which is the part i could check without it.
text only. images and files on the clipboard are ignored.
a 1-bit model can still be wrong in ways a number check can't see.

repo: https://github.com/walsoup/paste-sloth - python 3.11, MIT.

ClawbackMD: local Gemma drafts for AI-assisted insurance denials

Wal — Sat, 17 Oct 2026 23:07:49 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge: Week 2 — Side Hustle

What I Built

ClawbackMD drafts appeal packets for insurance denials that look AI-written on the payer side. You pick a denied claim, it reads the synthetic chart + payer policy, and it returns a verdict, quote-backed evidence, an editable letter, and peer-to-peer talking points.

The model is optional. Default store is in-memory (no Docker). If Ollama/Gemma is down, cached human-reviewed drafts still demo the flow.

Every letter is a draft. The UI asks for separate clinical and billing review logs. That is a demo attestation, not an e-signature and not NPI-bound.

Who Would Pay

Hospital billing / RCM teams who already lose hours to AI-assisted denials. A local open-weight draft that only cites verbatim chart quotes is something a small billing shop could run on a laptop without sending PHI to a hosted API.

Demo

Repo: https://github.com/walsoup/clawbackmd
Live demo: https://clawbackmd.vercel.app

npm install
cp .env.example .env.local
npm run dev

Optional model:

ollama pull gemma3
ollama serve

Five synthetic claims load on first open. Generate a draft on CLM-001 (appeal) and CLM-005 (documentation gap — query attending).

How It Works

Open-weight path: Gemma via any OpenAI-compatible endpoint (lib/llm.ts), usually local Ollama.

Guardrails in lib/validate.ts:

Evidence and policy quotes must be verbatim substrings of the source note/policy.
Letter sentences with clinical atoms (labs, grades, durations, CPTs) must appear in validated evidence or claim facts; ungrounded sentences are stripped.
The LLM may cite payer policy codes only. CFR / statute cites are attached server-side from plan_type, not invented by the model.
uphold_denial means documentation gap — query the attending — not automatic write-off.

UI uses the Linen design system (linen surface, forest primary, hairlines).

Why Open Innovation Matters

PHI stays local. Chart text never has to leave the machine if you run Gemma on Ollama.
Cost. No per-token bill for a drafting loop billing staff would run all day.
Swappable model. Point LLM_BASE_URL / LLM_MODEL at another OpenAI-compatible endpoint without rewriting the app.
Fail closed on quotes. If the model invents a finding, validators drop it instead of shipping a confident fake.

Tradeoff: a weak local model needs the cached drafts and the quote scrubber. That is intentional for a demo that must not invent clinical facts.

Code

https://github.com/walsoup/clawbackmd

roamshot: photo side quests for walks, with open models running on the phone

Wal — Sun, 11 Oct 2026 11:10:53 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass

What I Built

roamshot is an Android app for walks with something to look for. you pick a walk (a park, your neighborhood, a forest trail, or whatever you type in) and it hands you 3 to 5 photo side quests. each one has an objective and a difficulty. you go find something that fits, take the photo, and a small vision model checks whether it matches. if it misses you can retry or skip.

completed photos go into a gallery of instant-photo frames. a quest is worth 10, 20 or 30 points depending on difficulty, plus 5 for getting it on the first try, and one verified quest a day keeps your streak alive.

the point is that the phone is only there to give you a reason to look up. the quests point you at things in the world, and the answer is outside. there are no accounts, no backend, no social feed and no maps. it's for anyone who walks anyway and wants to pay a bit more attention.

quests are not allowed to ask for photos of people or faces, trespassing, private property, roads or traffic, unsafe night activities, or purchases. night walks, yes. horror-movie decisions, no.

Demo

no video demo.

i ran the whole thing on a low end budget phone, the worst one i have, on purpose. i also tested it offline, and the models downloaded and loaded on it, so the quests were generated by Qwen and not pulled from the built-in pool.

Code

walsoup / roamshot

Polaroam 📸

An indie Android walking game with on-device AI side quests and instant photo checks.
Crafted with ❤️ by @walsoup • Version 1.0.0

Polaroam turns everyday walks into retro photo scavenger hunts. Pick a walk style (or type your own), get creative on-device AI side quests, snap what you find through a pixel viewfinder, and pin your developed Polaroids to your corkboard gallery.

📱 Screenshots

Walk Setup	Quest Board	Camera Viewfinder	Polaroid Gallery	Stats & AI Config

Pick presets or custom walks	AI-generated side quests	Viewfinder with pixel reticle	Corkboard with Polaroids	Streak, score & @walsoup credits

✨ Features

🚶 Curated & Custom Walks: Choose from 7 themed presets (Neighborhood, City park, Forest trail, Coastal walk, Rainy stroll, Mountain trail, Night walk) or type whatever walk you're taking.
🎯 On-Device AI Quests: Generates 3–5 tailored photo objectives per walk with difficulty tiers (1–3 pips) and…

View on GitHub

Kotlin, Jetpack Compose, CameraX, Room, Hilt. the code lives in the repo under the package app.polaroam, which is the name the app had before it was roamshot.

How I Built It

every AI piece is an open model and runs on the phone.

job	model	runtime
writing quests	Qwen3-0.6B (Apache-2.0)	LiteRT-LM (Apache-2.0)
checking photos	CLIP ViT-B/32 (MIT), int8 ONNX export	ONNX Runtime (MIT)

quests. Qwen3-0.6B gets a system prompt and a walk type and writes the quests as JSON. the output goes through a parser and a safety filter. if it fails, the app retries once, and if that fails too it falls back to a built-in pool of canned quests, so a walk never starts with an empty board. there's also a smaller option, SmolLM2-360M, if the phone struggles with Qwen.

photo check. CLIP runs zero-shot. a script (tools/clip/build_label_embeddings.py) precomputes a text embedding for every label the quests can use, plus a few distractors like person, car and food, so an off-target photo has somewhere to land. on the phone, only CLIP's image encoder runs. a photo counts as a match when the quest's target labels together reach 20% and beat every other single label. that second rule is what stops a sandwich passing as a mushroom.

models are not in the APK. the Stats screen downloads both on Wi-Fi, about 700 MB with the default pair. the downloads are pinned to a specific Hugging Face commit and checked with sha256. until they're there, quests come from the built-in pool and photos can't be checked.

swapping models. the generator and the verifier are two interfaces (QuestGenerator, PhotoVerifier), and the choice of implementation lives in one file, di/AiModule.kt. you can point local.properties at a different .litertlm or ONNX file and rebuild. a different CLIP needs its own label embeddings.

Why Does Open Innovation Matter?

photos stay on the phone. the thing being sent to a model is a picture of wherever you are. with CLIP running locally, nothing you photograph leaves the device.
it works with no signal. after the one-time download, a walk needs no connection, and i tested it offline. trails are where signal is bad.
it costs nothing to run. there's no API bill, so no backend, no accounts and no limit on how many quests anyone generates.
I could swap models. the first plan was different (I looked at Gemma through LiteRT, which was slow to load on my phone, and at Moondream2). because the weights are open and the runtimes are open, trying Qwen3 and CLIP instead was a change to a config and a few classes, not a new vendor.

the tradeoff is honest: a 0.6B model writes plainer quests than a hosted one would, and CLIP zero-shot gets some photos wrong. that is the price of running on a phone, and for a "go look at a tree" app I think it's fine.