Inspiration
Taking electronics apart is easier than putting them back together. Suddenly you're wondering where everything belonged.
Manuals aren't always available, especially for older devices. Every device that becomes too frustrating to repair potentially ends up as another piece of e-waste.
I wanted to explore a simple idea:
What if your workbench could remember what you took apart and help you put it back together?
That idea became Anastylos, my solo project for the Meta AI Hackathon 2026.
What it does
Anastylos is an experimental repair assistant that combines computer vision, spatial evidence and conversational AI.
The goal is to turn a physical disassembly into a documented memory that can be consulted during reassembly.
The prototype brings together several capabilities:
- Camera capture: An Android APK records visual evidence during disassembly.
- Spatial calibration: Printed Aruco markers provide a measurable reference frame.
- Change detection: CV identifies changes inbetween observed states.
- Assembly Memory: Components, observations and user confirmations are saved with origin.
- Component drawings: Image segmentation and computer vision help extract component outlines.
- Exploded views: Documented components can be arranged into a SVG-based assembly representation.
- Conversational assistance: Meta Muse asks clarifying questions and helps retrieve information about the recorded components.
An important distinction: Anastylos does not autonomously understand every step of disassembly. The current MVP relies on operator confirmations for every part...
The AI suggests and the human confirms.
How I built it
Anastylos was developed primarily in Python, with a native Android APK written in Kotlin using Jetpack Compose and CameraX.
OpenCV handles calibration, image processing and change detection. Meta SAM 3.1 was used for segmentation, while Muse Spark provides conversational reasoning and clarification.
The Android application communicates with an authenticated receiver on desktop and the captured evidence can be associated with the project's assembly memory workflow.
The architecture separates observations from confirmed facts. A detected change doesn't automatically mean a component was removed and a plausible AI response isn't automatically accepted as truth.
I also experimented with microscopy, specialized component recognizers and geometric reconstruction. These remain separate, experimental parts of the project rather than fully validated features of the pipeline.
Challenges I ran into
Physical object tracking is difficult af
A component can disappear underneath a hand, changes its orientation, or become visually indistinguishable from another part. Tracking pixels is not the same as proving physical identity.
Increasing the video sampling rate produced more motion tracks but not necessarily better component identification.
Segmentation presented another problem: SAM sometimes included the objects shadow as part of its outline. Negative prompts removed the shadow but could also damage the actual outline.
A hybrid approach using SAM and computer vision for shadow subtraction produced better results in our experiments.
There were plenty of other difficulties: camera calibration, lighting, asynchronous communication, natural language interpretation and preserving assembly relationships between different stages.
Most of the exploded views looked like they had been drawn by a toddler.
Accomplishments that I'm proud of
The biggest milestone was completing a working, operator assisted workflow on an unfamiliar object.
During one test, Anastylos documented four confirmed components and four screws, preserved the recorded information and generated an exploded view representation.
We also demonstrated the interaction between visual evidence, Muse-powered clarification, assembly memory and component drawings.
The result was far from perfect but it brought previously independent systems together into a concrete prototype.
I am particularly proud of the work on evidence provenance and uncertainty. Instead of forcing the AI to hallucinate an answer, Anastylos is designed to preserve ambiguity and to request confirmation.
What I learned
This was my first hackathon and probably one of the most intense projects I've undertaken yet - It was really frustrating for most of the time but I learned a lot along the way.
The biggest lesson was that more capable AI models cannot compensate for observations that were never captured in the first place.
The right solution wasn't another model or a more sophisticated algorithm. It's a better camera angle, improved lighting or a well timed question to the operator.
I also learned a lot about computer vision, Android development, multimodal AI integration, hardware validation and coordinating complex development under a deadline.
Most importantly I learned the value of documenting failures rather than hiding them.
What's next
Anastylos remains an experimental MVP, not a production ready repair tool.
There is still substantial work to do: more reliable physical component identification, simpler calibration, improved drawings, stronger evidence replay and more useful reassembly guidance.
One particularly interesting direction is active perception: instead of passively watching the Agent could recognize missing evidence and ask the user to show a specific component or camera angle.
Whether or not Anastylos eventually becomes a finished product, the project has already opened up some fascinating questions about how AI can understand and interact with the physical world.
Built With
- android
- aruco
- camerax
- computer-vision
- git
- https
- javascript
- jetpack-compose
- kotlin
- meta-model-api
- meta-muse
- opencv
- python
- raspberry-pi
- rest-api
- sam-3.1
- svg
- typescript
