-
-
Math I did to come up with all thresholds with trials
-
Physical Bin
Binlytic AI
Binlytic is an AI-powered waste-management platform that identifies waste, validates where it belongs, physically routes it, verifies the disposal, learns from unknown objects, and turns every confirmed event into actionable data.
Instead of treating a trash bin as the end of the waste process, Binlytic turns it into part of a connected waste-intelligence network.
$$ \text{Identify} \rightarrow \text{Validate} \rightarrow \text{Route} \rightarrow \text{Verify} \rightarrow \text{Learn}, repeat. $$
Inspiration
Waste sorting sounds simple until you're standing in front of three bins holding a coffee cup, food container, or half-empty bottle and trying to figure out where it belongs.
Overflowing and poorly managed waste can also create unsanitary conditions, attract pests, produce odors, and increase people's exposure to waste in shared spaces. Better sorting and faster response therefore isn't only an efficiency problem. It contributes to cleaner, safer, and more usable public environments as Poorly managed waste can expose workers and the public to broken glass, decomposing organic matter, pests, mold, and other potentially hazardous conditions. By improving sorting and helping facilities respond to problems earlier, Binlytic aims to contribute to cleaner and safer waste-management environments, particularly for the workers who interact with waste every day
We started with one question:
What if people didn't have to decide where their waste belonged and helped workers and officials make garbage management safer and efficient?
That grew into a larger system designed not only to sort waste, but also to verify what actually happened, learn from new objects, monitor infrastructure, and connect citizens with the organizations responsible for waste management.
What it does
A camera observes an item placed in front of Binlytic.
OpenCV processes the frame and CLIP compares the object against waste descriptions in a shared embedding space.
For an image vector V and text vector T:
$$ s(I,T)=\frac{v_I \cdot v_T} {\lVert v_I\rVert \lVert v_T\rVert} $$
But Binlytic does not simply take the highest-scoring result.
Before anything moves, the prediction passes through:
- minimum similarity rejection
- destination-margin checks
- contamination conflict checks
- framing checks
- confirmation across consecutive frames
Our current minimum similarity boundary is approximately (0.20). This is a cosine-similarity threshold, not 20% confidence.
If Binlytic isn't sufficiently certain, it returns UNKNOWN instead of forcing a bad classification. We currently have a garbage compartment fallback for that as we dont want the person to not have a place where we can throw trash so after a 7 second timeout, if no successful classification - it goes into garbage.
Once accepted, local disposal rules determine whether the object belongs in recycling, compost, garbage, or another supported stream.
The destination is sent to an ESP32, which controls the physical routing servos.
But that still isn't enough.
Prediction does not equal confirmation.
An ultrasonic sensor waits for the item to actually enter the selected compartment. Only then is the disposal recorded.
$$ \text{Camera} \rightarrow \text{AI} \rightarrow \text{Decision} \rightarrow \text{Servo} \rightarrow \text{Sensor} \rightarrow \text{Confirmed Event} $$
In our documented prototype testing, Binlytic achieved approximately 92% end-to-end success, where success required both correct routing and completion of the physical disposal flow.
More than a smart bin
Every confirmed disposal becomes part of the Binlytic operations platform.
Organizations such as malls, campuses, schools, municipalities, and commercial facilities can monitor connected units, confirmed disposal activity, estimated stream capacity, collection forecasts, contamination events, alerts, system health, employees, locations, and waste trends.
Employees also have a mobile interface for viewing assigned bins, priority alerts, network status, and issues requiring service.
Binlytic therefore doesn't stop after sorting an object.
It turns physical disposal events into operational intelligence.
Citizen reporting
Binlytic hardware can see what happens around its own installations.
Citizens can see everything else.
The citizen side of the platform allows people to report problems such as litter, overflowing bins, illegal dumping, damaged infrastructure, missed collection, and contamination.
Reports can include a photo and location and can be routed to a configured organization or municipal contact.
Other citizens can vote on existing reports, helping repeated community observations gain greater visibility.
$$ \text{Smart Bin Data}+\text{Citizen Reports} \rightarrow \text{Waste Intelligence} $$
Binlytic learns
Unknown objects are not simply discarded from the system.
Binlytic supports both manual learning and automatic evidence-based learning.
A human can review an unknown object, label it, and approve its CLIP embedding as a learned reference.
Repeated unknown objects can also form clusters. Once enough evidence is collected, including observations across separate installations, they can enter the shared learning process.
Approved references are stored online and synchronized across connected Binlytic units.
$$ \text{One Binlytic Learns} \rightarrow \text{Network Synchronizes} \rightarrow \text{Other Units Benefit} $$
Recognition knowledge can be shared while local disposal rules remain specific to each location.
How we built it
Binlytic combines computer vision, AI, embedded hardware, IoT, web software, and mobile software.
The main pipeline is:
$$ \text{Webcam} \rightarrow \text{OpenCV} \rightarrow \text{CLIP} \rightarrow \text{Validation} \rightarrow \text{Local Rules} \rightarrow \text{ESP32} \rightarrow \text{Servo} \rightarrow \text{Ultrasonic Verification} \rightarrow \text{Dashboard} $$
The computer-vision pipeline is written primarily in Python using PyTorch, Hugging Face Transformers, OpenCV, NumPy, and CLIP ViT-B/32.
An ESP32 handles the physical layer, including routing servos, ultrasonic sensors, camera positioning, environmental sensing, OLED output, device health, and network communication.
The web platform provides the organization and citizen experiences, while the employee mobile interface provides an operational workflow for staff.
The challenge wasn't simply getting each technology to work.
It was getting all of them to behave as one closed-loop system.
Challenges we ran into
Knowing when AI should not act
CLIP will almost always produce a closest result, even for an unfamiliar image. We had to build rejection thresholds, margins, stability checks, and conflict logic so Binlytic could recognize uncertainty instead of blindly trusting an argmax.
Prediction wasn't proof
One of our most important design changes came when we realized that logging an AI classification immediately would create data about what the model thought should happen, not what physically happened.
We redesigned the pipeline around ultrasonic confirmation.
Now the AI prediction and the completed disposal are separate states, and an event becomes confirmed only after the hardware detects the physical disposal.
Connecting AI to hardware
A correct classification on a laptop means little if the wrong servo moves, the object never falls, or a sensor fails to register it.
Synchronizing inference, networking, servo movement, sensor timing, and dashboard state required significant testing.
Real-world vision
Hands, backgrounds, changing lighting, reflections, orientation, and partially visible waste all affect recognition. We added framing and background-processing logic to make classification more robust.
Accomplishments that we're proud of
We're most proud that Binlytic became a functioning physical/software ecosystem rather than an image-classification demo.
We built the complete path:
camera → AI → validation → routing → physical movement → sensor verification → dashboard
Alongside that core system, we built:
- physical three-stream routing
- local CLIP inference
- uncertainty rejection and multi-frame verification
- contamination handling
- ultrasonic disposal confirmation
- organization operations dashboard
- employee mobile workflow
- citizen reporting and community voting
- collection and capacity intelligence
- unknown-object learning
- shared learning between connected units
The prototype was built using inexpensive, accessible components, supporting our longer-term goal of retrofitting existing waste infrastructure instead of replacing every bin.
What we learned
The biggest lesson was:
$$ \text{AI Prediction} \neq \text{Engineering Certainty} $$
A model producing an answer doesn't mean a physical system should immediately act on it.
We learned to design uncertainty, verification, hardware failures, network problems, and unknown objects into the architecture itself.
We also learned that the most valuable data isn't necessarily what the AI thinks happened.
It's what the system can verify actually happened.
And perhaps most importantly:
A smart system shouldn't just make decisions. It should know when to question them.
What's next for Binlytic AI
Next we want to expand real-world testing across more environments, waste types, lighting conditions, and users.
We're also exploring:
- load cells for direct weight measurements
- improved capacity sensing
- expanded network learning
- larger multi-location deployments
- more advanced organization analytics
- deeper municipal reporting integrations
- Binlytic Mobile, a future mobile waste station for temporary high-demand areas
The long-term value of Binlytic isn't one intelligent bin.
It's what becomes possible when an entire network of them can sort, verify, learn, and share intelligence together.
Log in or sign up for Devpost to join the conversation.