← Selected work

On-device applied ML · 2025

EcoGuardian AI

The useful model is the one that fits the device, responds in the moment, and still clears an evidence-based quality bar.

Source ↗

82%

model accuracy

<300ms

on-device inference

−60%

model size

01

Problem

Waste-sorting guidance is most useful at the moment of disposal, but a cloud-dependent classifier adds latency, connectivity risk, and privacy cost to a simple decision.

02

Why it matters

A phone-based system can make guidance immediate and accessible. Running locally also keeps camera input on device and makes the core classification flow available offline.

03

Architecture

  1. 01Camera frame
  2. 02Image preprocessing
  3. 03Quantized TFLite model
  4. 04Material class
  5. 05Disposal guidance
  6. 06Optional Appwrite sync
The classification path stays on device. Cloud services are separated for optional identity, persistence, and community features rather than required for inference.

04

Technical challenges

01

Fitting the model to the phone

Used post-training quantization to shrink the model while monitoring accuracy and latency together.

02

Designing for the camera loop

Structured the prototype around capture, review, classification, and clear disposal guidance instead of a generic ML result screen.

03

Keeping offline as the default

Separated inference from optional backend features so the core utility does not disappear with connectivity.

05

Tradeoffs

Quantized model over maximum desktop accuracy

Sub-300ms local feedback was part of the product requirement, not a later optimization.

Three actionable categories

Recycling, compost, and landfill map model output to a decision people can make immediately.

Optional cloud services

Community and history features can sync later without making core classification fragile.

06

Experiments

  1. 01Measured accuracy after post-training quantization.
  2. 02Benchmarked inference on the target mobile path rather than extrapolating from desktop performance.
  3. 03Tested the product flow around a camera-first interaction and immediate guidance.

07

Results

Reached 82% classification accuracy in the project benchmark.

Reduced model size by 60% while keeping inference under 300ms.

Delivered a fully offline core classification experience.

08

Lessons learned

  • Latency and package size are model-quality dimensions on edge devices.
  • A constrained label space can produce a clearer product decision than a larger but ambiguous ontology.
  • Cloud independence is both a reliability and privacy feature.

09

Future work

  • Validate on a broader, more representative image set.
  • Complete end-to-end inference wiring and field testing across target devices.
  • Add calibrated uncertainty and a safe fallback when the image is out of distribution.