DEV Community: Ankita

I Built a Ghost That Shows You Your Future as a Comic Book (100% on My Laptop, Wi-Fi Off) 🎃🕯️

Ankita — Sun, 01 Nov 2026 05:56:06 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine

On Día de los Muertos, we build an ofrenda for the people we miss. I wondered what it would look like to build one for a future that hasn't happened yet.

What I Built

Ofrenda of Maybe turns a wish into a comic book, entirely on my own laptop.

You whisper a future you hope for ("I have a big garden full of sunflowers and a small cottage in the hills"), and the app gives you:

📖 A 4-panel comic page. Three panels tell your wish as a cheerful cartoon story.
👻 A ghost panel. The ghost is you, from that future. It remembers the very first small step that started it all.
🕯️ The ghost's offering: one small, real thing you can do today. My favorite so far: "Buy one sunflower seed today."
🎬 A narrated film of the same story, with four narrator voices (including a ghostly whisper).
🎃 A Halloween UI: cobwebs, a dangling spider, a flickering title, and a ghost that may jump out while you type (you can switch it off in Settings).

I took the theme two ways at once. Halloween is the spooky wrapper (the ghost, the webs, the jump scare). Día de los Muertos is the heart: an offering, a way of remembering, but pointed at the future instead of the past. The "ghost in the machine" here is a friendly one, a future you living inside a model on your own GPU.

Demo

Everything in the video runs on one laptop with the Wi-Fi turned off.

Code

Ankita562 / ofrenda-of-maybe

Describe a future you hope for. A friendly ghost turns it into a candlelit animated story, entirely on your own laptop with open models.

🎃 Ofrenda of Maybe

Whisper a future you hope for. A friendly ghost from that future turns it into a short cartoon film, entirely on your own laptop.

Built for the Hacktoberfest Open-Source AI Challenge, Week 4: Ghost in the Machine.

How it works

flowchart LR
    subgraph L["💻 Your laptop (no cloud, works offline)"]
        direction LR
        A["💭 Your wish"] --> B["🧠 Gemma<br/>writes 4 scenes"]
        B --> C["🎨 SD-Turbo<br/>draws them"]
        C --> D["🎙️ Piper<br/>narrates"]
        D --> E["🎬 Cartoon film"]
        B --> F["🕯️ The ghost's offering:<br/>one small first step"]
    end

Scenes 1-3 tell your wish as a cheerful story. Scene 4 is the ghost: your future self, who remembers the first small step that started it all. That step also appears as its own card, so a daydream becomes something you can do today.

Features

🎬 4-scene cartoon film with captions and slow zoom
🕯️ "The ghost's offering": one…

View on GitHub

How I Built It

 💭 your wish
    │
    ▼
 🧠 Gemma 3 (4B, via Ollama) ── writes 4 scenes + the "first small step" (JSON)
    │
    ▼
 🎨 SD-Turbo (diffusers) ────── draws one cartoon picture per scene
    │              │
    ▼              ▼
 🎙️ Piper TTS     📖 Pillow ─── lays out the comic page (captions, speech bubble)
    │
    ▼
 🎬 MoviePy ─────────────────── narrated film

Gemma 3 4B (through Ollama) writes the storyboard. I force a JSON schema, so it always returns exactly the fields I need: scenes, narration, image prompts and the first step.
SD-Turbo (🤗 diffusers) paints each scene. Gemma is unloaded from the GPU before the image model runs, which is how the whole thing fits in 4 GB of VRAM (RTX 3050 Laptop).
Piper gives the narration, locally, with a voice picker.
Pillow draws the comic page and MoviePy builds the film. Gradio is the UI.

What broke along the way

A version mismatch. AttributeError: module 'torch' has no attribute 'accelerator'. The newest diffusers wanted a newer PyTorch than the CUDA 12.1 build I had installed. Moving to the cu126 build fixed it.
My own prompt leaked. I had written "the bakery, the pastries" as examples in the system prompt. A 4B model copied them, so my famous singer ended up performing in a bakery. Deleting the examples fixed it.
Crowds became blobs. SD-Turbo can't draw an audience, so "cheering crowd" turned into round circles, and my first "ghost" was just a round blob. I now tell Gemma to avoid crowds, faces and hands, and I describe the ghost as "a cartoon ghost made of a white bedsheet".
The ghost had no point. Version one just repeated the story. I changed it into your future self, who remembers the first small step, and gave that step its own card.

Why Does Open Innovation Matter?

Your hopes stay yours. People type very personal wishes. With local models they never touch a server.
It works with the Wi-Fi off. I filmed the demo that way. No API key, no account.
It's free to run, and the pieces are swappable: the storyboard model is one line (LLM = "gemma3:4b"), the art style is a prompt string, and new narrator voices are just downloaded files.
Control over memory. On a 4 GB GPU, being able to load and unload models myself is what made this possible at all. A closed API wouldn't let me.

The Spirit of It

My favorite moment came from the sunflower wish. The ghost didn't repeat the story back to me. It said, "I remember when you planted that first seed," and handed me a step I could take that same day: "Buy one sunflower seed today."

That's the idea behind an ofrenda: an offering that keeps something alive. This one just keeps a future alive.

Your turn: what future would you ask the ghost about? Drop a wish in the comments and I'll run it and reply with the ghost's first small step. 🕯️

Prize Categories

🔮 Best Use of Gemma: Gemma 3 4B runs locally through Ollama and writes the storyboard, the narration lines and the first small step as structured JSON.
🔊 Best Use of ElevenLabs: the voice-over in my demo video was generated with ElevenLabs. (The app itself narrates with local Piper.)

I Built an Offline Face Finder So My Friends Can Stop Scrolling the Group Chat.

Ankita — Sun, 04 Oct 2026 14:46:15 +0000

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend

What I Built

We took a one-day trip to Mangalore, and the moment we got back, the photos started flooding the group chat. My friend let's name her Poo wanted to post a few on Instagram, but first she had to scroll through all of our pictures to find the ones she was in. It took about two hours.

That's when I thought: what if something could do this for us? Finding "the photos with my face in them" is a job a computer is good at. But it would mean handing my friends' faces to someone else's server, and I wasn't willing to do that. So I built FaceGotcha for Poo and the rest of our group: a version that runs entirely on your own laptop and finds your photos for you.

FaceGotcha is a "where am I in all these photos?" machine. You give it a few photos of your face and a giant pile of group photos, and it hands back only the ones you're in. Nothing is uploaded anywhere.

The whole thing in five steps:

Pick up to 5 clear photos of yourself.
Throw in the photo pile: a folder, like an unzipped WhatsApp export.
Press Find my photos (the app cheerfully announces that it's "ignoring photos of food...").
Swipe through only the uncertain ones. Swipe right for "that's me" (cue the GOTCHA! stamp) and left for "nope". It's like a dating app, except every match is you.
Download a zip of your photos.

In my test of 55 WhatsApp photos, 26 were matched automatically, 9 needed my opinion, and 20 were ignored. Your original photos are never touched.

Faces are biometric data, and these are my friends' faces. So the one rule was: nothing leaves the laptop. No accounts, no cloud, no API keys.

Demo

Code

Ankita562 / facegotcha

Find the photos you're in from a group-chat photo dump. Fully local face matching with open models (OpenCV YuNet + SFace). No cloud, no uploads.

Find the photos you're in, from a group-chat photo dump, fully offline

Drop in a few photos of your face and a folder of group photos. FaceGotcha finds the ones you appear in, lets you swipe through the uncertain ones, and gives you a zip. It runs on your own computer with open-source face models, so nobody's face data is ever uploaded anywhere.

Built for the Hacktoberfest Weekend Challenge: Build for a Friend.

Demo

Video: Watch the demo video

Why I built it

One friend in my group has the phone with the best camera…

View on GitHub

How I Built It

Two small open models do the heavy lifting. Both come from the OpenCV Zoo and run on the CPU, so no GPU is needed:

YuNet finds every face in a photo.
SFace turns each face into a fingerprint of 128 numbers. Faces of the same person land close together, and faces of different people land far apart.
OpenCV runs them, and Flask serves a little web page at 127.0.0.1, so other devices on the network can't open it.

The scoring works like a bouncer at a door. A photo scoring 0.53 or higher walks straight in. Between 0.38 and 0.53, it gets an ID check, which means it comes to you for a swipe. Below 0.38, it's turned away.

I added that middle band because OpenCV's suggested single cut-off (0.363) let the wrong people in during my first test:

Setting	Auto-matched	Wrong person in it
OpenCV's default cut-off (0.363)	35	8
My cut-off with a review band (0.53 and 0.38)	26	0

The swipe review is where I spent the most design effort. It's one photo and one gesture at a time, with undo on one key (Z). A "nope" doesn't delete anything: the photo moves to a Rejected album you can browse or download. The page also has a progress tracker along the top that ticks off each step, a scrolling reel of your matches, and a stats bar showing how every photo was sorted. Confetti appears when you find your photos, because finding your photos should feel like winning.

I built this with AI assistance (Claude) for writing and explaining the code. I designed the idea, tested it on my own photos, and tuned the cut-offs from the scores.

Honest limits:

This is a small test on one person's face, not a benchmark, and the cut-offs were tuned on that one sample.
It missed one photo of me, a group shot that scored 0.176. No cut-off could catch it without flooding the review pile with wrong people.
Small faces in compressed group shots are the hardest case.
A photo counts as a match if any face in it matches, which is why the review step matters.
The stats bar shows how photos were sorted, not an accuracy score. The app has no answer key, so I counted the real accuracy by hand.

What's next: finding photos of two people at once (me and my best friend, together and solo), and collecting every photo of a friend for a birthday post.

Why Does Open Innovation Matter?

A face-sorting app has a problem most apps don't: it has to look at every one of your friends' faces. Sending the whole group chat to a cloud service just to find my face would be like photocopying everyone's face to find mine, including people who never agreed to that.

Open models made a different design possible:

Privacy: the models run on your own machine, so face data never leaves it.
Cost: no API keys, rate limits, or per-photo charges. The only download is two small model files, and after that it works offline.
Control: I could see what the models do, change the cut-offs, and swap in a different open model without asking anyone.
Reuse: both models are open source (YuNet is MIT-licensed and SFace is Apache 2.0, according to their LICENSE files in OpenCV Zoo), so anyone can run, change, and build on this project.

The honest trade-off: a big cloud service would probably be more accurate out of the box. Mine needs reference photos, a tuned cut-off, and a human to settle the borderline cases, and for photos of my friends, I'll take that deal.

DEV Community: Ankita

I Built a Ghost That Shows You Your Future as a Comic Book (100% on My Laptop, Wi-Fi Off) 🎃🕯️

What I Built

Demo

Code

Ankita562 / ofrenda-of-maybe

Describe a future you hope for. A friendly ghost turns it into a candlelit animated story, entirely on your own laptop with open models.

🎃 Ofrenda of Maybe

How it works

Features

How I Built It

What broke along the way

Why Does Open Innovation Matter?

The Spirit of It

Prize Categories

I Built an Offline Face Finder So My Friends Can Stop Scrolling the Group Chat.

What I Built

Demo

Code

Ankita562 / facegotcha

Find the photos you're in from a group-chat photo dump. Fully local face matching with open models (OpenCV YuNet + SFace). No cloud, no uploads.

Find the photos you're in, from a group-chat photo dump, fully offline

Contents

Demo

Why I built it

How I Built It

Why Does Open Innovation Matter?