Accenture scales video evaluation with Amazon Nova and Amazon Bedrock Brokers

This publish was written with Ilan Geller, Kamal Mannar, Debasmita Ghosh, and Nakul Aggarwal of Accenture.

Video highlights provide a strong strategy to enhance viewers engagement and lengthen content material worth for content material publishers. These quick, high-impact clips seize key moments that drive viewer retention, amplify attain throughout social media, reinforce model identification, and open new avenues for monetization. Nonetheless, conventional spotlight creation workflows are sluggish and labor-intensive. Editors should manually evaluate footage, establish vital moments, minimize clips, and add transitions or narration—adopted by guide high quality checks and formatting for distribution. Though this offers editorial management, it creates bottlenecks that don’t scale effectively.

This publish showcases how Accenture Highlight delivers a scalable, cost-effective video spotlight era resolution utilizing Amazon Nova and Amazon Bedrock Brokers. Amazon Nova basis fashions (FMs) ship frontier intelligence and industry-leading price-performance. With Highlight, content material house owners can configure AI fashions and brokers to help numerous use circumstances throughout the media {industry} whereas providing a human-in-the-loop choice for high quality assurance and collaborative refinement. This maintains accuracy, editorial oversight, and alignment with model pointers—with out compromising on velocity or scalability.

Actual-world use circumstances

Highlight has been utilized throughout a variety of {industry} eventualities, together with:

• Personalised short-form video era – Highlight’s specialised brokers analyze well-liked short-form content material (equivalent to video reels and different social media) to establish patterns of high-performing content material. The brokers then apply this understanding to long-form video to generate personalised quick clips, with built-in checks for model alignment and content material requirements.
• Sports activities modifying and highlights – Highlight automates creation of video highlights for sports activities like soccer, Components 1, and rugby, tailoring them to particular person preferences and pursuits. It additionally validates every spotlight’s high quality and accuracy, streamlining editorial workflows in consequence.
• Content material matching for stakeholders – Utilizing enriched metadata, Highlight matches archived or reside video content material to viewers demographics, optimizing distribution methods and maximizing advertiser worth via exact focusing on.
• Actual-time retail provide era – In retail environments equivalent to gasoline stations, Highlight processes reside CCTV footage to deduce buyer profiles utilizing information (equivalent to automobile sort or transaction historical past), after which dynamically generates personalised product affords. These affords think about contextual elements equivalent to time of day and climate; and they’re delivered with customized visuals in close to actual time.

Highlight’s structure

Highlight’s structure addresses the problem of scalable video processing, effectively analyzing and producing content material whereas sustaining velocity and high quality. It incorporates each task-specific fashions and Amazon Nova FMs which are orchestrated by specialised Amazon Bedrock brokers. Key architectural highlights embrace:

• Job-driven mannequin choice – Highlight dynamically selects between conventional AI fashions and Amazon Nova FMs based mostly on a given job’s complexity and latency necessities. This clever orchestration permits quick inference for time-sensitive operations whereas deploying deeper multimodal reasoning the place refined evaluation is required—balancing velocity and intelligence throughout functions from real-time retail affords to advanced video processing.
• Agent orchestration – Specialised brokers, every purpose-built for particular evaluation duties, function throughout the end-to-end workflow beneath the course of a central orchestrator agent. The orchestrator agent manages job breakdown, information circulation, and inter-agent communication.
• Scalable and adaptable – Through the use of AWS capabilities, Highlight’s structure is configurable to help completely different workloads—from high-throughput video spotlight era to low-latency provide personalization on the edge.

Highlight makes use of a multi-layered agent workflow to automate video processing and era whereas sustaining high quality management. For instance, to generate dynamic video highlights, Highlight makes use of three specialised “tremendous brokers” that work in coordination beneath a central orchestrator agent’s supervision. Every tremendous agent is powered by Amazon Nova fashions, and is supported by a set of utility brokers (see the next diagram). These brokers work collectively to know video content material, generate high-quality highlights, and keep alignment with person necessities and model requirements.

The workflow consists of the next tremendous brokers and utility brokers:

• Video processing agent – This agent analyzes long-form video and generates detailed metadata to information short-form video creation. It makes use of the next utility brokers:
  • Analysis agent – Analyzes well-liked short-form movies to establish key elements that create video virality, and creates recipes for profitable short-form content material. For instance, in music movies, it may well spotlight choreographed dance sequences with the lead performer as important segments and a recipe based mostly on this perception.
  • Visible evaluation agent – Applies the analysis agent’s findings to new long-form content material. It identifies matching segments, tags key people, and timestamps related moments. It makes use of conventional AI fashions (equivalent to individual recognition and monitoring) to seize fine-grained particulars for phase identification.
  • Audio evaluation agent – Performs speech diarization and transcription to help each the analysis and visible evaluation brokers with deeper context from the video’s audio observe.
• Brief video era agent – This agent orchestrates the precise creation of the short-form video by integrating related segments and refining the sequence. Its utility brokers embrace:
  • Part of curiosity (SOI) agent – Identifies potential segments based mostly on video style, goal size, featured performers, and JSON metadata from the visible evaluation agent. This agent prioritizes logical circulation and viewer engagement.
  • Video era agent – Constructs video utilizing phase suggestions and part patterns from the video processing agent. For instance, influencer movies would possibly comply with a construction of an attention-grabbing hook, key messages, and a name to motion. The method might be iteratively improved based mostly on suggestions from the reviewer agent.
  • Video postprocessing agent – Refines the ultimate output for publishing by performing duties like cropping to mobile-friendly side ratios, or including subtitles, background music, and model overlays.
• Reviewer agent – This agent works iteratively with the era agent to take care of video high quality and relevance. Its utility brokers embrace:
  • Relevance test agent – Evaluates alignment with user-defined content material pointers, viewers expectations, and desired themes.
  • Abruptness test agent – Offers clean transitions between segments to keep away from jarring cuts, enhancing viewer expertise and professionalism.

See Highlight in motion:

Answer overview

To work together with Highlight, customers entry a frontend UI the place they supply pure language enter to specify their goal. Highlight then employs its agentic workflow powered by Amazon Nova to attain its given job. The next diagram illustrates the answer structure for video spotlight era.

The workflow consists of the next key elements (as numbered within the previous diagram):

1. Frontend UI for person interplay:
   1. Customers work together via an internet portal secured by Amazon Cognito authentication and delivered utilizing Amazon CloudFront.
   2. Amazon API Gateway serves a restful endpoint for video processing and spotlight era companies.
2. Dwell video stream processing:
   1. AWS Elemental MediaLive processes incoming video stream and triggers AWS Lambda to provoke workflows. (Highlight additionally accepts video archive content material as media information for processing and spotlight era.)
3. Video processing workflow orchestrated with AWS Step Features:
   1. Open supply fashions hosted on Amazon SageMaker allow speech evaluation and laptop imaginative and prescient for individual and object detection.
   2. The video processing agent powered by Amazon Nova Professional analyzes video and generates fine-grained metadata (for instance, figuring out patterns from viral movies).
   3. The reviewer agent powered by Amazon Nova Premier maintains alignment with model requirements.
   4. Open supply utility tooling is used for pre-analysis duties.
4. Spotlight era workflow orchestrated with Step Features:
   1. Amazon Nova Professional analyzes the person question for clips of curiosity to know intent, and reformulates the question for downstream processing.
   2. The quick video era agent powered by Amazon Nova Professional constructs a video spotlight utilizing phase suggestions.
   3. The reviewer agent powered by Amazon Nova Premier makes certain the constructed spotlight aligns with high quality, model, and contextual expectations.
   4. AWS Elemental Media Convert and open supply tooling allow video spotlight building and postprocessing (equivalent to subtitle layover, side ratio change, and transitions).
5. Storage and monitoring:
   1. Amazon Easy Storage Service (Amazon S3) shops metadata extracted from processing workflows, reference content material (equivalent to scripts and model pointers), and generated outputs.
   2. Amazon CloudWatch maintains end-to-end system well being and displays efficiency.

Key advantages

Highlight’s method to video processing and era creates dynamic worth. Moreover, its technical design utilizing Amazon Nova and an built-in agentic workflow helps content material house owners notice beneficial properties of their video processing and editorial operations. Key advantages for Highlight embrace:

• Cross-industry utility – Highlight’s modular design permits it to be utilized seamlessly throughout industries—from media and leisure to retail
• Actual-time processing – It helps each reside stream feeds and pre-recorded video, with customized spotlight era occurring in minutes, decreasing from hours or days
• Value-efficient deployment – It’s fully serverless and on-demand, minimizing idle infrastructure prices and maximizing utilization
• Effectivity – Accenture’s evaluate of prices utilizing Amazon Nova fashions confirmed that Amazon Nova-powered brokers ship over 10 occasions higher value financial savings over conventional spotlight creation strategies

The next desk offers is a comparative evaluation of Highlight’s video processing method to standard approaches for video spotlight creation.

Conclusion

Highlight represents a cutting-edge agentic resolution designed to sort out advanced media processing and buyer personalization challenges utilizing generative AI. With modular, multi-agent workflows constructed on Amazon Nova, Highlight seamlessly permits dynamic short-form video era. The answer’s core framework can be extensible to numerous {industry} use circumstances that require multimodal content material evaluation at scale.

As an AWS Premier Tier Providers Accomplice and Managed Providers Supplier (MSP), Accenture brings deep cloud and {industry} experience. Accenture and AWS have labored collectively for greater than a decade to assist organizations notice worth from their functions and information. Accenture brings its {industry} understanding and generative AI specialists to construct and adapt generative AI options to consumer wants. Along with AWS, via the Accenture AWS Enterprise Group (AABG), we assist enterprises unlock enterprise worth by quickly scaling generative AI options tailor-made to their wants—driving innovation and transformation within the cloud.

Check out Highlight on your personal use case, and share your suggestions within the feedback.

In regards to the authors

Ilan Geller is a Managing Director within the Knowledge and AI apply at Accenture. He’s the World AWS Accomplice Lead for Knowledge and AI and the Middle for Superior AI. His roles at Accenture have primarily been targeted on the design, improvement, and supply of advanced information, AI/ML, and most just lately Generative AI options.

Dr. Kamal Mannar is a World Pc Imaginative and prescient Lead at Accenture’s Middle for Superior AI, with over 20 years of expertise making use of AI throughout industries like agriculture, healthcare, vitality, and telecom. He has led large-scale AI transformations, constructed scalable GenAI and laptop imaginative and prescient options, and holds 10+ patents in areas together with deep studying, wearable AI, and imaginative and prescient transformers. Beforehand, he headed AI at Vulcan AI, driving cutting-edge innovation in precision agriculture. Kamal holds a Ph.D. in Industrial & Techniques Engineering from the College of Wisconsin–Madison.

Debasmita Ghosh is working as Affiliate Director in Accenture with 21 years of expertise in Info Know-how (8 years in AI/Gen AI functionality), who presently amongst a number of duties leads Pc Imaginative and prescient apply in India. She has offered her paper on Handwritten Textual content Recognition in a number of conferences together with MCPR 2020, GHCI 2020. She has patent granted on Handwritten Textual content Recognition resolution and acquired recognition from Accenture beneath the Accenture Inventor Award Program being named as an inventor on a granted patent. She has a number of papers on Pc Visions options like Desk Extraction together with non-uniform and borderless tables accepted and offered within the ComPE 2021 and CCVPR 2021 worldwide conferences. She has managed initiatives throughout a number of applied sciences (Oracle Apps, SAP). As a programmer, she has labored throughout numerous phases of SDLC with expertise on Oracle Apps Growth throughout CRM, Procurement, Receivables, SCM, SAP Skilled Providers, SAP CRM. Debasmita holds M.Sc. in Statistics from Calcutta College.

Nakul Aggarwal is a Topic Matter Knowledgeable in Pc Imaginative and prescient and Generative AI at Accenture, with round 7 years of expertise in growing and delivering cutting-edge options throughout laptop imaginative and prescient, multimodal AI, and agentic methods. He holds a Grasp’s diploma from the Indian Institute of Know-how (IIT) Delhi and has authored a number of analysis papers offered at worldwide conferences. He holds two patents in AI and presently leads a number of initiatives targeted on multimodal and agentic AI. Past technical supply, he performs a key position in mentoring groups and driving innovation by bridging superior analysis with real-world enterprise functions.

Aramide Kehinde is World Accomplice Options Architect for Amazon Nova at AWS. She works with excessive progress firms to construct and ship ahead pondering expertise options utilizing AWS Generative AI. Her expertise spans a number of industries, together with Media & Leisure, Monetary Providers, and Healthcare. Aramide enjoys constructing the intersection of AI and artistic arenas and spending time along with her household.

Rajdeep Banerjee is a Senior Accomplice Options Architect at AWS serving to strategic companions and shoppers within the AWS cloud migration and digital transformation journey. Rajdeep focuses on working with companions to offer technical steering on AWS, collaborate with them to know their technical necessities, and designing options to satisfy their particular wants. He’s a member of Serverless technical subject group. Rajdeep is predicated out of Richmond, Virginia.