The starting point
- Training videos were searched for on platforms by hand, watched and labelled in Excel.
- Duplicate videos and inconsistent labels made the data messy.
- Outreach to vet clinics, animal shelters and influencers ran through spreadsheets and inboxes.
Data pipeline: WoofCollector
We built a web app that covers the whole path from video to labelled training file:
- Search: find videos automatically on YouTube, Instagram and through Google search.
- Download and cut: into short clips for training.
- Detect duplicates: using audio fingerprints.
- Label: in a dedicated interface instead of Excel.
- Query the model: send clips to the startup’s AI model and store the predictions.
- Store: sync clips and metadata to cloud storage automatically.
Result: labelling throughput went from 5 to 30 videos per hour, 6 times more, with better label quality.
Sales pipeline: Zoho CRM
- Workflows for outreach to vet clinics, animal shelters and influencers: qualification, routing, statuses, follow-ups and reminders
- Data flows via webhooks into a PostgreSQL database
- Funnel and conversion metrics in Metabase dashboards and automated Slack reports to management
Infrastructure and compliance
- Self-hosted PostgreSQL and n8n on a cloud VM with Docker
- GDPR consent management and documentation with the EU AI Act in mind from the start
- Handover documentation with risks and next steps
Who on the team worked on it
Process analysis, CRM, data pipeline and project lead: Margarita. Web app development: Dima.