AI
Ingestion Without Data Pipelines
Coactive is a model-agnostic platform. Feed it images or video from Azure Blob Storage or any cloud storage and it generates embeddings using open-source foundation models, closed-source models, or models from Azure AI Foundry. No data pipelines, no ML operations to configure.
The ingestion API handles all of that. Developers point to storage, Coactive ingests the content and makes it queryable within seconds. The web UI exists for demos; most customers work directly through the API.
Semantic Search on Visual Content
Basic semantic search works out of the box. Search for “explosions” or “interview with Jim from the office” and the embeddings return matching frames. That part requires nothing beyond what a foundation model already knows.
Enterprise use cases break that. Lowe’s needed to identify employees wearing a Lowe’s vest. No foundation model is trained on “Lowe’s associate.” Generic embeddings alone cannot close that gap.
Zero-Shot Classification with a Feedback Loop
Dynamic tags convert text prompts into zero-shot classifiers across an entire dataset. In the dynamic tags live demo (04:39), Komala typed five categories, including “fight scenes” and “cinematic landscape,” and Coactive scored every asset immediately. The model flagged Paramount logo frames as cinematic landscape because of the sunset and snow-capped mountains.
“A double click is a negative prompt. A single click is a positive prompt.” – Satya Komala
Clicking those frames as negatives retrains the classifier in real time. Adding the text prompt “greenery” shifted results instantly. Fandom uses the same mechanism for trust and safety, scoring user uploads against its own content standards as they arrive.
SQL Queries Over Unstructured Video
Every tagged dataset gets a SQL representation automatically. Customers write standard T-SQL against tables for dynamic tags, audio, and transcripts. One demo query showed that “action sequences” drove the most clicks and “vehicle shots” drove the most impressions, joining visual metadata with engagement data.
The Mimir MAM integration demo (12:15) made the gap concrete: keyword search returned nothing for “flight taking off,” while Coactive returned the exact moments within videos where planes lift off. Samsung Reuters uses this for its editorial asset library. A forthcoming auto-evaluator generates LLM-based ground truth and surfaces F1 and recall scores before teams train downstream models.
Notable Quotes
A double click is a negative prompt. Satya Komala · ▶ 6:22
A single click is a positive prompt. Satya Komala · ▶ 6:24
using visual AI with Coactive is extremely easy. Satya Komala · ▶ 14:17
Key Takeaways
- Point Coactive at Azure Blob or any cloud storage and embeddings generate without writing data pipeline code.
- Dynamic tags use text prompts plus click-based feedback to retrain visual classifiers in real time.
- A SQL layer lets analysts query image and video metadata the same way they query structured tables.