Reeloli

India's local-business owners already know how to sell. What they often do not have is a content strategist, copywriter, music producer, editor and campaign manager available every day. Most AI creative tools still ask them to learn a new interface, write prompts and manage a collection of disconnected outputs.

Reeloli turns the channel they already use—WhatsApp—into an AI-operated content studio. A business owner can click an Instagram ad, open WhatsApp, and explain the business naturally through messages, product photos, voice notes, menus or brochures. Reeloli converts that conversation into an evidence-backed business brief, asks only the next useful question, surfaces contradictions, and never treats an AI inference as a confirmed business fact.

From one conversation to an entire content calendar

Once the owner confirms the context, Reeloli plans a usable 15-day content calendar: one hero video, three lightweight calendar videos and four image posts, each with a date, format, hook, message, call to action, approved asset references and exact source facts. The planner rejects invented prices, guarantees, rankings or performance claims and refuses a plan with missing, duplicate or out-of-window posts.

This is not a chat response that disappears when the tab closes. Reeloli creates a durable background job and starts the planning work asynchronously. The owner can leave WhatsApp while the system packs the request, submits it to a Batch provider, polls it later, validates the strict eight-post result, renders an infographic and PDF, and places both in an operator-review queue. Every transition is stored in Cloudflare D1, request and result files are kept in R2, and each provider operation has an idempotency key, budget ceiling, usage record and recovery path. Only an approved plan is delivered back to the customer with clear next-step choices.

We have proven this loop in a controlled live UAT. One confirmed bakery context became a valid 15-day plan through a real background Batch job using 496 input tokens and 1,660 output tokens at a reconciled provider cost of ₹2.61. The first render exposed clipped Day 13 and Day 15 cards, so delivery stopped. We repaired the renderer, visually verified the complete infographic and two-page PDF, sent both over WhatsApp, and recorded the customer's interactive request for the content program. That is the product philosophy in miniature: queued is not done, generated is not approved, and paid work is never silently repeated.

How Gemini and Lyria power the studio

Our Gemini stack is deliberately layered. Gemini 3.5 Flash-Lite is the high-throughput multimodal intake model: it turns messy, multilingual business material into structured evidence and helps choose the next question. Gemini 3.1 Pro acts as a bounded creative supervisor. It does not rewrite an entire campaign; it either accepts a plan or returns the smallest typed repair for a specific story, script or production defect. This separation gives us strong reasoning without handing a probabilistic model control of facts, payments or retries.

Lyria 3 Clip is our audio-first creative engine. Reeloli develops three distinct 30-second directions before visual production, so the customer can listen, choose and lock the emotional idea first. Lyria produces the musical or vocal performance; the approved audio is then immutable. Visuals, captions and motion must follow its measured timeline, instead of rewriting the soundtrack after the customer has approved it. In our current live content-program canary, three real hero auditions were generated, technically validated and delivered in WhatsApp; the hero video remains correctly paused until the customer chooses.

Our next Gemini expansion is asynchronous visual inspection with Gemini 3.6 Flash through the Gemini Batch API. Google currently documents Gemini 3.6 Flash as its production model for fast multimodal and agentic work; we therefore use that official name rather than an unofficial "3.7 Flash" catalog label. The planned inspector will review contact sheets and rendered frames for brand fidelity, readable text, continuity, framing and unsafe claims before an operator sees them. Batch mode fits Reeloli because these checks are important but not latency-sensitive, and Google documents it as an asynchronous route for large request volumes at reduced cost. This capability is roadmap, not a claim about the current demo.

A reliable AI business, not just a generator

Reeloli runs on Cloudflare Workers, Durable Objects, Queues, D1, R2, Browser Rendering and a wake-on-demand media container using FFmpeg and FFprobe. Deterministic code owns identity, confirmed facts, money, budgets, leases, idempotency, artifact hashes and delivery. AI owns the work where judgment creates leverage: context interpretation, content planning, creative routes, audio direction, visual planning and bounded repair. The owner confirms facts, chooses the audio and approves the preview; an operator resolves ambiguous paid provider outcomes.

The system now goes beyond planning. An accepted content program materializes eight independently scheduled executions. The live canary has delivered its first deterministic image post, moved the hero into customer audio choice, kept three future images ready for their dates, and blocked three montage videos until approved customer media exists. No model is allowed to invent the missing assets, and no future delivery is presented as completed.

What we learned and what comes next

The hardest problem was not making one impressive image. It was preserving truth and customer intent across a long-running, multi-provider business. We learned that an affordable AI agency needs inspectable state, small repair nodes, audio-first approval, immutable assets and human authority at the exact points where reputation or money can be harmed.

Next we will complete direct Gemini API observability, add the Gemini 3.6 Flash Batch visual-inspection gate, finish the customer-approved hero and montage canaries, and measure time-to-calendar, approval rate, delivered cost, revision rate and enquiries generated across a small founding cohort. Our goal is not to replace the local business owner's voice. It is to give that voice a dependable creative operation that can keep working after the conversation ends.

Built With

• cloudflare-queues
• cloudflare-workers
• d1
• durable-objects
• ffmpeg
• gemini-3.1-pro
• gemini-3.5-flash-lite
• gemini-3.6-flash
• gemini-api
• google-ai-studio
• lyria-3
• openai-batch
• openrouter
• r2
• typescript
• whatsapp-cloud-api