Edge AI
In progress
6 classes
SubScan
Subsurface pest detection from seismic and acoustic signals
Detecting pests underground by listening to them: a seismic and acoustic signal pipeline meant to classify six pest types from vibration signatures on a microcontroller in the field. Currently at the synthetic-data stage.
- Working synthetic-data pipeline in Python: FFT features (100 bins, 0–500 Hz) into a 1D CNN, 6 classes. Class separation is artificial at this stage, so no accuracy figure is claimed.
- Next: hardware integration (STM32F411, ADXL345 accelerometer, microphone) and field data collection in Oct 2026.
- Planned: post-training quantization to fit a sub-100 ms inference budget on constrained hardware.
PythonKerasNumPySciPySTM32Signal processing
github.com/eliftb/SubScan ↗
Computer Vision
Benchmark
4 models
Video Action Recognition
SlowFast vs VideoMAE vs TimeSformer vs I3D
Runs one video through four pretrained Kinetics-400 models and compares what each predicts, how confident it is, how fast it runs and how much memory it needs.
- Measured on an RTX 4060 laptop GPU: I3D takes 22.7 ms per clip end to end, VideoMAE 102.3 ms, about 4× slower for the highest published accuracy.
- Published accuracy figures are labelled as not measured here;
evaluate.py measures top-1 and top-5 on your own labelled videos.
- Modular package (decoding, preprocessing, model wrappers, benchmark, visualisation) with tests that run without model weights. Outputs a summary figure and an annotated video.
PythonPyTorchHugging FacePyTorchVideoOpenCV
github.com/eliftb/video-action-recognition ↗
Systems
Sample data
30/30 planted errors
Payout Reconciliation
Payout audit for restaurants on Turkish delivery platforms
Computes what a delivery platform should have paid from the restaurant's own POS records and contract terms, compares it with the payout statement, and attributes every missing lira to exactly one cause.
- Date-ranged commission rules (rate, base, fixed fee, processing fee, VAT, delivery) with a line-by-line breakdown for every order.
- 12 discrepancy types; each underpaid amount gets a single cause, so the dispute total is never double-counted.
- On the generated benchmark month: all 30 planted errors found, 0 false positives, a 3,599.95 TL shortfall (5.4%) fully explained. Not yet run on a real platform statement.
- Python standard library and SQLite, runs locally; unit, regression and HTTP tests with GitHub Actions CI.
PythonSQLiteopenpyxlGitHub ActionsTesting
github.com/eliftb/tr-payout-reconciliation ↗
NLP
Research
90.2% accuracy
Turkish Emotion Classifier
7-class NLP system with a measured negative result
A local Turkish emotion classifier, and an honest account of the one class the data cannot support.
- Assembled a 7-class training set from three public Turkish datasets plus my own hand labels, using a purpose-built annotation interface; TF-IDF plus logistic regression, running locally with no API dependency.
- 90.2% accuracy on a stratified, held-out 7,135-record test set.
- Published a Cohen's kappa inter-annotator analysis quantifying the data ceiling for the rare class — contempt, 13 samples, 0.167 F1 — and documented it as a data constraint, not a model one.
- Flask labeling UI with live agreement metrics, plus a Chrome extension that scores page text with the local model.
Pythonscikit-learnFlaskChrome ExtensionAnnotation
github.com/eliftb/tr-emotion-class ↗
Backend
Course project
22 tests
Hasad
Digital marketplace backend, producer to consumer
A full-stack Spring Boot and Next.js marketplace connecting local producers with consumers, built for a software engineering course: layered REST API, role-based access, migrations and Docker.
- JWT authentication with four roles (consumer, producer, admin, super admin); orders, cart, reviews, producer approval and an admin panel.
- PostgreSQL schema managed with Flyway migrations; Swagger UI for the API.
- JUnit and Mockito unit and integration tests for the auth, order, product and producer services (22 tests).
- Dockerized backend and database with a PostgreSQL health check, via docker-compose.
Java 17Spring Boot 3.2PostgreSQLFlywayDockerSpring Security
github.com/eliftb/digital-marketplace-hasad ↗
Computer Vision
Shipped
Claude Vision
Drape
AI virtual outfit try-on
A try-on app that reads the garment and the body separately, then puts them together: vision-model analysis for the clothing, pose landmarks for the fit.
- Claude Vision API for garment analysis, MediaPipe pose landmarks for AR fitting.
- Emotion detection layer feeding outfit recommendation.
- Full JavaScript client with the vision pipeline running end to end.
JavaScriptClaude APIMediaPipeComputer Vision
github.com/eliftb/drape ↗