I previously built Labellerr, a data‑annotation platform for image, audio, and video workflows, with a focus on human‑in‑the‑loop automation.
Labellerr was created to support multimodal annotation (image, audio, video) and to tightly integrate human‑in‑the‑loop automation, allowing model predictions to pre‑populate labels that annotators can accept, modify, or reject. I chose a custom build over off‑the‑shelf tools to embed specific automation loops and to scale annotation pipelines for machine‑learning projects.
- Key decisions: custom platform for deep automation integration; prioritized support for three media types; implemented human‑in‑the‑loop loops to balance speed and quality.
- Methods: web‑based UI for labeling; model inference APIs that pre‑populate labels; feedback loop retrains models from annotator corrections; task‑management assigns work by load and expertise.
- Open questions: which automation techniques (active learning, model‑assisted labeling) were used? how was throughput measured and what gains were observed? what scaling strategies (distributed workloads, cloud infra) were employed? are there published benchmarks showing impact on model accuracy or labeling cost?