Case Study: Building a Modular AI Influencer System — Generated and Posted to Social Media Without Human Touch

Client: Jure Erlic — AI content business operator (Canada) Engagement: Fixed-price, 3 weeks (Jan 23 – Mar 17, 2026) Verified Upwork contract — 5.0 / 5.0 mutual rating

The brief (in the client's words)

Jure didn't come to me with vague pain — he came to me with a clean spec:

"Design and implement a modular, automation-first AI influencer system capable of generating and posting realistic image content to social media at scale, while minimizing detection risk through controlled content variation, behavioral randomization, and perceptual realism. The system must be: Scalable — easy to add new houses and influencers. VA-friendly — operable by non-technical staff. Future-proof & Modular — supports new image models, formats, and platforms. Risk-aware — optimized for longevity and reach, not overengineering."

Translated: he wanted an AI content engine that runs without him, scales to multiple personas, can be operated by a non-technical VA, and is built to last as the AI tooling landscape evolves.

The problem

Before this build, Jure was running every step of the content pipeline manually:

Brainstorming concepts, outfits, locations, and poses for each post
Writing prompts for the image model
Generating images through Wavespeed
Stripping AI metadata from the output files
Uploading and scheduling each post

The bottleneck wasn't image generation — that's fast. The bottleneck was creative direction (figuring out what to generate) and mechanical operations (everything between "image done" and "post live"). Producing daily content across multiple AI personas was unsustainable for one person.

What I built: The AI Influencer Swarm Manager

A modular, multi-persona AI content pipeline. The architectural decision I'm proudest of is the "Houses and Rooms" concept — and it's worth explaining because it's what made the personas feel like real Instagram identities instead of random AI output.

The "Houses and Rooms" Architecture Rather than generating images in random environments, I designed each AI persona's world like an actual home. The base has a Houses table and a Rooms table — some rooms are shared across personas (kitchen, living room, balcony) and some are private to each persona (bedroom, dressing room, personal workspace). Every generated image is anchored to a specific room. Followers subconsciously recognize the spaces over time, which is exactly how real influencer accounts build visual identity.

The Pipeline (end to end)

Auto-Prompt Generation — From a small set of keywords per persona, GPT generates high-quality, detailed image prompts handling styling, clothing, pose, lighting, and room context. Eliminates the creative bottleneck.
Scheduled Triggering — At a configured time of day per persona, Airtable kicks off the generation run.
AWS Lambda Webhook — Airtable sends the record ID to a Lambda function, which pulls full record details via the Airtable API.
Image Generation — Lambda passes the prompt and persona reference images to Wavespeed (Flux Schnell, Flux Pro, and Google Nano-Banana 2 models).
Postback Lambda — A second Lambda receives the generated image, runs a Python script to strip all AI metadata, then organizes and uploads the file to AWS S3 in the correct persona/room folder structure.
URL Attached Back to Airtable — Only the S3 URL gets stored, not the file itself. (Important cost decision — see below.)
Posting — An n8n workflow picks up the ready post and publishes it on schedule.

Tech stack: Airtable (Houses, Rooms, Influencers, Variation Seeds, Generation Jobs, Posting Logs, Blotato Accounts), AWS Lambda, AWS S3, Wavespeed (Flux Schnell, Flux Pro, Google Nano-Banana 2), GPT, n8n, custom Python.

A specific cost-engineering decision worth calling out: Storing AI-generated images directly in Airtable attachments would have ballooned the storage bill quickly — Airtable charges for attachment volume above plan limits. By keeping images in S3 and only storing URLs in Airtable, the system stays cheap to run regardless of generation volume. Same logic I apply on every project: architect for the bill, not just the feature.

VA-Friendly by Design Per the brief, the system had to be operable by non-technical staff. Adding a new persona or new room is a configuration task in Airtable — no code changes, no AWS console. A virtual assistant can run day-to-day operations without engineering involvement.

Build time: 3 weeks, fixed-price contract.

The outcome

~35+ hours per week saved. Jure is functionally not involved in day-to-day operations of this content stream anymore.
Daily posting consistency — algorithms reward consistent posting; the system delivers it without human attention.
Multi-persona scalability — adding a new AI influencer is a configuration task, not a build task. The architecture is true to the brief.
Pay-per-use cost profile — AWS Lambda + S3 means the client pays only for what runs.
Still in production — the system has generated thousands of images and is running today.

What the client said

Verified Upwork review, March 2026 — 5.0 / 5.0:

"Shehar is really great to work with. He is excellent at understanding client needs and executing the plan as agreed upon. I highly recommend working with him! I'd love to work together again in the future." — Jure Erlic, Canada

What this means for you

If you're running an AI content operation — virtual influencers, brand personas, automated social accounts — there's a clean architecture that takes the human out of every step except strategy. Airtable as the brain. AWS as the muscle. n8n as the publishing arm. Built right, the marginal cost of adding another persona is configuration, not engineering.

If you're running a marketing agency thinking about offering AI content services to your clients — the same backbone applies. Build it once, deliver it as a service.

[Want a system like this for your business or your clients? → Book a free 20-minute audit call]