Selected work / Case study
SafetyAI
A web workspace for reviewing construction helmet predictions, connecting a Next.js interface to FastAPI and a fine-tuned YOLO26s model. Supports annotated results and browser-local inspection summaries.

The problem
Give users a review workflow for submitted site images and videos, with visible predictions, downloadable results, and inspection summaries separate from illustrative demo records.
Engineering focus
Connect media selection, explicit inference, original/annotated comparison, and local result summaries; distinguish the online review application from the offline model-training workflow.
Implementation overview
Browser workspace
Select media and explicitly start analysis in the Next.js interface.
Inference API
The browser uploads directly to the configured FastAPI service.
Helmet detector
A fine-tuned YOLO26s checkpoint produces class predictions and image annotations.
Human review
Compare media, inspect class counts, and download results.
Local summaries
The browser stores image-summary metadata, not full image or video payloads.
Implementation
- Build a bilingual Next.js workspace for overview, detection, and model guidance, with explicit analysis actions and original/annotated image review.
- Connect the browser directly to a configured FastAPI inference service that loads the trained checkpoint and returns class counts and annotated output.
- Organize dataset inspection, pretrained YOLO26s fine-tuning, evaluation, and optional export in an offline Python workflow; preserve the saved baseline curves as historical evidence.
- Offer media downloads and browser-local image-summary history, with search, filters, and CSV export; keep full images/videos out of persisted history.
- Implement a video-processing path using temporary files and an optional codec-conversion step; real video/codec execution was not verified in the supplied review.
Evidence & source
The package includes a recorded image-inference response and a saved 50-epoch baseline plot. Training-validation and separate evaluation scores differ and are not merged into one performance claim. The plot and recorded result were not reproduced during this portfolio update.
Decisions & limitations
Predictions require human review; class counts overlap and are not unique worker counts. The recorded image output has a blue cast and overlapping labels. Video/codec behavior and field reliability remain unverified. The dashboard uses illustrative demo statistics; saved training-validation scores differ from separate evaluation records. No continuous CCTV ingestion, multi-user database, or authentication service is implemented.