Ai.Fish

Domain-first product split: They expose an API for integration into third-party EM (electronic monitoring) review systems while running their own cloud web app (CatchVision). This implies a two-pronged engineering approach: lightweight, standardized inference endpoints plus a more feature-rich human-in-the-loop review UX — not just a single monolithic product.

Edge-aware engineering implied by founder background: the founder’s Control Engineering / edge-hardware focus strongly suggests they are optimizing models for onboard/low-bandwidth environments (quantized models, low-power inference, prioritized upload), even though the site doesn’t explicitly say so. That’s an unusual emphasis compared with startups only offering cloud inference.

Video-first, multi-task CV pipeline is implied (object detection + species classification + event segmentation/counting). Handling these together requires specialized architectures (temporal models, instance tracking, multi-head networks) rather than naive frame-by-frame detectors — a technical step up from many ‘image-only’ CV plays.

Human-in-the-loop active learning workflow is implied by wording about reducing reviewer work to confirming annotations. This suggests they have a feedback loop: model proposes annotations, humans validate, labels flow back to model retraining — a production data loop that’s essential but often missing in early-stage CV startups.

Operational-scale video pipeline and storage optimizations are a core unstated challenge. Targeting 10M+ hours/year implies they’ve had to design chunked ingestion, event-level indexing, cheap cold storage + hot caches, and metadata-first retrieval (not typical for small CV teams).