Making AI feel
calm and useful

I’m John, a designer and developer in Manila. I build AI products that work where people need them: Ligaya, a voice agent for emergencies, and Lumina, a family of image, video, and science tools that run entirely on a laptop GPU.

  1. fig. 01 · Ligayaandroid
    Ligaya app home screen with the Talk to Ligaya button and SOS
    1. say “Ligaya, tulong” to start a call
    2. shares your location with responders
    3. one-tap emergency, no talking needed

    Android · An emergency voice agent. Hears “Ligaya, tulong,” asks what’s wrong, and alerts the nearest police, fire, or medical service.

    Visit Ligaya
fig. 05 · Lumina Type● live tester
Lumina Sans · 400 · 46px31 chars
click the row again to closefull specimen →
Ligaya

An emergency agent that
listens, locates, and calls

Visit Ligaya

Ligaya listens for distress shouts like “Ligaya, tulong,” finds the nearest police, fire, and medical services, and talks you through it with two-way speech. A one-tap SOS button means mute and non-verbal users can call for help too.

Ligaya, tulong!
I’m here. What’s happening?
There’s smoke on the second floor.
I’m alerting the nearest fire station now. Stay low and head for the stairs.
Calling the fire station
Voice firstSpeech-to-text and text-to-speech run in a loop, so Ligaya can listen and answer while your hands are busy.
Distance from youSample view

Fire station alerted1.2 km away

SOS

Can’t speak? One tap sends the same alert.

Station dispatchPinpoints the closest emergency services so help is routed quickly.

Wake phrase

Listens for shouts like “Ligaya, tulong” in Filipino.

One-tap SOS

Built for mute and non-verbal users who can’t speak up.

Android · Kotlin · Gemini

A native Android app with Gemini doing the reasoning.

Lumina

Image and video generation
that never leaves your laptop

How it works

Two local-first studios built to run open models on an 8 GB laptop GPU. Prompts, images, and videos stay on the machine, and the servers only listen on 127.0.0.1. No accounts, no uploads, no cloud bill.

What should we draw?

4 imagesSquareAny styleAvoid…
Animagine XL · ≈ 6 min per run
Create
Lumina output: a girl with a warm drink by a rainy café windowGenerated
Lumina output: a café at night lit by paper lanterns
Lumina output: anime-style cherry blossoms against a blue sky
Lumina StudioDraw anime with Animagine XL 4.0 or photos with Krea 2 Turbo, then upscale or cut out the background. Write a sentence and Lumina turns it into tags.
00:00
1664×9608 stepsRIFE 32 fpsAniSora V3.1
She rests her cheek on her hand and slowly turns a page
Animate
Lumina VideoBring a picture, describe the motion, and get anime. Plan whole stories as shots, or wire every step together in a node editor.

Sentences to tags

Studio rewrites a plain prompt into the tags Animagine was trained on.

Upscale 4×

Tiled Real-ESRGAN, so VRAM stays low on large images.

Storyboard

Video plans a story as shots joined by crossfades, with a title and music.

Extend to 30 s

Each 5-second part starts from the last frame, so the character stays the same.

Lumina Science

A research workbench
that cites its sources

View on GitHub

A local science assistant that runs Qwen3.5-9B through Ollama on your own GPU. It searches more than 250 million papers, reads your PDFs, runs Python in a sandbox, and draws interactive visuals. Every claim is numbered back to a real source you can check.

Lumina Science on a phone: the graphs page with charts from experiments
GraphsEvery chart from your research lands in one gallery, drawn live from the data so you can hover and zoom.
3D moleculesProteinsDNA helixCell diagramsProcess diagramsMechanism of actionDose-response curvesReaction schemesPeriodic tableField linesWave interferenceProjectile motionPunnett squaresSIR epidemicsEnzyme kinetics3D surfacesSlope fieldsMatrix transformsDistributions
Interactive visualsThe model picks the visual that helps and fills it in, from 3D proteins to physics and maths with live sliders.

Deep research

Plans sub-questions, reads the literature, and writes a cited report.

Experiments

Hypothesis, method, code, results, and limitations, saved as a write-up.

Checked citations

Sources that don’t match a real paper are flagged as unverified.

Your library

Drop in PDFs or spreadsheets and ask about them, with page-level citations.

Under the hood

Fitting big models
into 8 GB of VRAM

More on GitHub

Lumina Studio

A FastAPI server, a no-build web UI, and a job queue that survives a restart. Animagine XL 4.0 fits fully in 8 GB; bigger models get a memory mode picked from your VRAM.

Animagine XL 4.0The default: 4 images per run in about 6 minutes on an RTX 5050 Laptop
Draft, then finishA quick preview, re-rendered at full quality with the same seed
2 MP and 4 MPRendered at 1 MP and built up, so memory never runs out
Sentences to tagsPlain prompts become the Danbooru tags the anime model understands
# Lumina Studio: local image generation .\setup.ps1 -Full # PyTorch (CUDA 12.8) + diffusers .\run.ps1 # opens http://127.0.0.1:7860 # models animagine-xl-4 SDXL anime, fits in 8 GB # default krea2-turbo 12B, 8 steps, no CFG # memory mode for Krea, picked from your VRAM 40 GB+ full GPU 20–40 GB model CPU offload 8–16 GB NF4 4-bit, toggle in Settings

Lumina Video

AniSora V3.1: Wan2.1-I2V-14B fine-tuned on anime and distilled to 8 steps. Its fp8 weights stream to the GPU one block at a time while the previous block computes, and the 5.7B text encoder runs on the CPU.

16.4B parameters, 8 GBAbout 0.8 GB of weights sit in VRAM at once; the rest waits in a memory-mapped file
5 seconds in 10 minutesDown from 36, with leaner blocks, SageAttention, and fp8 tensor-core matmuls
Up to 30 secondsBuilt from 5-second parts, each starting from the last frame of the one before
Only what changed re-runsStoryboard and node editor share one graph, cached node by node
# Lumina Video: one shot, from picture to library picture your image, or Animagine XL draws one ↓ prompt the motion, plus AniSora's motion score ↓ animate AniSora V3.1 · 8 steps · 5 s parts at 16 fps ↓ finish Real-ESRGAN to 720p · RIFE to 32 fps ↓ soundtrack MusicGen, or your own track ↓ library saved with its graph, reopens in Workflow
Contact

Have something
worth building?

I’m open to freelance work, internships, and conversations about AI tools that should feel simpler than they do. I usually reply within a day or two.

Email copied