CurrentCut

CurrentCut keeps company video as current as the product itself

Every company changes faster than its content calendar. A new product ships, an offer evolves, an event approaches, or a conversation starts trending — but the social feed is often still telling last quarter’s story. Marketing teams are then forced to choose between slow manual research and fast content that may be stale, generic, or disconnected from what the company actually does.

We were looking for a tool to help our social media team create reels that reflect the company’s current products and services. That search led to CurrentCut: an agentic content operation that keeps the feed alive by turning current, source-backed signals into reviewable video ideas. The marketing team reviews and approves concepts; CurrentCut then carries the approved concept through the production pipeline so the team can move from “what should we say now?” to a ready-to-review vertical reel.

The product is designed to support a go-to-market strategy, not replace marketing judgment. Its promise is simple: CurrentCut keeps company social video as current as the product itself.

Features and functionality

• Source-first company context. A company website is the canonical starting point. Optional business context, brand assets, and visual guidance help the agent represent the organization accurately.
• Background content scouting. Configurable schedules let the Content Agent revisit enabled sources and look for timely news, seasons and events, and social-trend or meme-format signals. A single bounded heartbeat expands due schedules into tenant-scoped asynchronous work instead of keeping a browser request open.
• Grounded research with evidence. URL Context establishes company facts; Google Search grounding supplements them when external corroboration is enabled. Signals carry citations, freshness, risk, and a “why now” explanation so a reviewer can see the reasoning behind a recommendation.
• Review-ready concept generation. Each concept includes a hook, proposition, target audience, treatment, opening beat, visual continuity rules, scene plan, caption, call to action, influence labels, citations, brand-fit status, and estimated cost. The agent can retain a GENERATE or SKIP decision and uses semantic evidence to reduce repetitive ideas.
• Human approval before spend. Concepts and opening frames are reviewable first. Video, voice, and music work is bound to the approved concept revision, keeping the expensive production path behind an explicit marketing decision.
• Vertical reel production. The approved path generates a 9:16 image-to-video asset with Veo, adds optional Gemini TTS narration and Lyria instrumental music, then finalizes the media with a deterministic renderer and immutable checksums. The result stays linked to its concept, source, prompt, model, approval, and asset lineage.
• Operational safety. Firestore records are tenant-scoped; provider credentials stay on the server; App Check and role checks protect control routes; event delivery is idempotent and retryable; budgets, safety outcomes, and failures are recorded. Automatic social posting is deliberately disabled so publishing remains a human-owned action.
• Editorial library. Approved concepts, generated assets, captions, evidence, versions, and finalization metadata remain searchable and recoverable for the team’s review workflow.

Technologies used

Google SDK and agent framework. The selected Google agent framework is the Google GenAI SDK (google-genai), used through @google/genai with Vertex AI. The worker uses the SDK for structured Gemini generation, URL Context and Google Search grounding, opening-frame generation, and Veo image-to-video operations. Supporting Google client libraries handle Pub/Sub, Cloud Storage, and Text-to-Speech calls. The repository does not use Genkit or Antigravity; ADK is not the selected framework for this submission.

Google AI models.

• Gemini 3.6 Flash — primary reasoning model for grounded research, brand and content strategy, campaign decisions, triage, structured concept generation, and critique. This satisfies the hackathon’s Gemini 3.5-or-newer requirement.
• Gemini 3.1 Flash Image — opening frames and visual anchors for concepts.
• Veo 3.1 (veo-3.1-generate-001) — approved-concept, first-frame-guided vertical video generation.
• Lyria 3 Pro Preview (lyria-3-pro-preview) — original instrumental background music.
• Gemini 2.5 Flash TTS (gemini-2.5-flash-tts) — optional narration using the configured Sulafat voice.

Application and infrastructure. CurrentCut is a TypeScript monorepo with a React/Vite web studio, Firebase Authentication and App Check, Cloud Run services, Firestore, Pub/Sub, Eventarc, Cloud Storage, a controlled Cloud Scheduler heartbeat, and FFmpeg-based deterministic media finalization. Versioned Zod contracts connect the browser, control API, scanner boundaries, agent worker, and renderer without putting provider keys in browser code.

Other data sources used

• The organization’s public company website, fetched as the canonical source through URL Context and bounded source-policy controls.
• Public web results retrieved through Gemini’s Google Search grounding when a schedule or run has enabled news, event, or trend research.
• Optional organization-provided business context and rights-attested brand/reference assets.
• Internal run state in tenant-scoped Firestore: schedules, source evidence metadata, citations, decisions, approvals, concepts, media lineage, usage, and recovery records.

CurrentCut treats retrieved web pages as untrusted evidence, not instructions. A social-network OAuth connector and automatic posting action are not enabled in this build; the output is a reviewable, export-ready asset and the team owns the final publishing decision.

Findings and learnings

1. “Current” has to be provable. A trend keyword alone is not enough. A useful marketing idea needs a company fact, a fresh signal, a citation, and a clear explanation of why the idea is timely.
2. The valuable agent action is the workflow, not the prose. The strongest result is a durable chain from scheduled research to a deduplicated decision, reviewable concept, approved media asset, and searchable lineage.
3. Asynchronous state changes the product experience. Claims, leases, inbox/outbox records, retries, and recovery state let the team refresh or return later without losing work or accidentally starting duplicate paid generations.
4. Human approval is a feature. It protects brand accuracy, keeps budgets predictable, and gives marketers control over which timely ideas become real content.
5. Multimodal production needs separate failure boundaries. Keeping research, opening frames, video, narration, music, and finalization as distinct jobs makes each result auditable and lets a failed later step retry without regenerating every upstream asset.
6. Safety and GTM speed reinforce each other. Source restrictions, tenant isolation, model attribution, cost ceilings, and no-posting defaults make an autonomous system easier for a marketing team to trust and adopt.

Built With

• cloud-run
• cloud-scheduler
• cloud-storage
• eventarc
• ffmpeg
• firebase
• firestore
• gemini-2.5-flash-tts
• gemini-3.1-flash-image
• gemini-3.6-flash
• google-genai-sdk
• google-search-grounding
• lyria
• marketing-automation
• multimodal-ai
• pub/sub
• react
• typescript
• url-context
• veo-3.1
• vertex-ai
• vite