2,500 species in the library.
Not one of them was searchable.

We didn't pick a model and hope. The first proof of concept (POC) processed 10,000 minutes of wildlife video. It sampled key frames from each clip and used Amazon Nova models on Bedrock to identify species, behavior, habitat and technical attributes:

• The pipeline generated up to ten metadata fields per clip, linked to its video ID in Amazon OpenSearch.
• Video and query representations supported semantic retrieval alongside the structured metadata.
• A Streamlit interface let users describe a request, retrieve candidate clips and inspect the results.
• Candidate approaches, including Nova and Twelve Labs, ran against the same footage, so the model decision was made on comparative results.

Why: we ran the models side by side on the library itself, because benchmarks on general images say little about wildlife footage.

A later AWS Experience-Based Acceleration tested species identification on a bounded set of 20 species, at production quality, and set out the work needed to extend coverage to the full catalog. It also tested containerized transcoding against real RED and ProRes files, and delivered deployment runbooks and reusable implementation patterns.

A buyer's brief mixes the creative with the technical: what's happening, where, and at what frame rate and format. The AI-agent search design handles both:

• Bedrock parses the request into species, behavior, context and technical requirements.
• Semantic retrieval finds relevant content, while structured filters apply requirements such as frame rate and format, and results are ranked for visual and behavioral relevance.
• A batch processor accepts spreadsheets or task lists from research buyers, runs every requested search and assembles the matching material into organized collections for review.
• Metadata generation describes interactions, setting and cinematic attributes as searchable tags and editorial summaries, reaching tagging accuracy above 95% across species, behavior, habitat and technical attributes, as the company reports.

Temporal analysis opened a new way to sell the footage. It divides behavior into sequences of roughly five seconds, each with a behavior label, a confidence score and context. Scientific institutions and AI companies want behavioral data at that grain, not whole clips, so the library gained a dataset-licensing business alongside conventional clip sales.

EventBridge and Step Functions connect the private media archive to the licensing workflow, coordinating quality assessment, enrichment, licensing readiness and publication, so moving a clip from the archive to the marketplace no longer waits for a person to trigger it.

Better search is worthless if footage can't get onto the platform in a usable form. We replaced polling with event-driven processing:

• S3 multipart upload, Transfer Acceleration and presigned ingestion support resumable submission of large professional files, through a cross-platform command-line upload tool, with single files up to 10 TB across more than 50 video formats.
• Input-format detection selects the transcoding path, with REDline and FFmpeg for decoding and conversion and g4dn and g4ad GPU instances for acceleration.
• Step Functions tracks each job. Recoverable failures retry with exponential backoff, and unrecoverable jobs move to a dead-letter queue to be investigated.
• Each upload produces seven preview assets, meaning five video-quality previews, a still image and a proxy for AI processing and editorial review.
• CloudWatch shows job progress and failures. Spot capacity and autoscaling hold down compute cost, with GPU autoscaling responding in under three minutes, and S3 Intelligent-Tiering follows changing storage access.

When a step fails, the pipeline retries it and then parks the job where someone will see it, so no footage is lost.

The architecture assessment mapped the whole existing stack, a legacy PHP CodeIgniter backend and a WordPress frontend, and defined the target for four platforms: private media asset management, stock-footage licensing, AI-dataset licensing and video transcoding. The target combines containerized services on ECS, Aurora MySQL for relational data, DynamoDB for workflow state, OpenSearch for retrieval and AWS Batch with CPU/GPU workers, with a JavaScript frontend on S3 and CloudFront replacing the PHP and WordPress pattern.

AWS Marketplace readiness was designed in from the start, with metering, entitlement and billing integration for subscriptions and private offers, rather than added later. The security stack (AWS WAF, Shield, CloudTrail and KMS) is aimed at SOC 2 Type II and GDPR readiness for enterprise buyers.

A dedicated DevOps pod looked after the live service in parallel. It investigated abnormal traffic, tuned WAF rules to block bot and malicious requests, fixed the database load they caused, and improved logs and alerts so the team sees a problem before users do.