Most of our work at Partenit lives close to the metal: edge AI, retrieval-augmented generation, systems built under the EU Horizon dAIEDGE umbrella. This article is about what happened when we pointed that stack at a deliberately unglamorous problem: small business social media. The result is Laspi, an AI social media content generator that turns a two-minute voice note into a week of ready-to-publish content, and the engineering behind it turned out to be more interesting than we expected.
The problem: content is a second job nobody applied for
Every small business owner has heard the same advice: post consistently, know what to post on Instagram for your business, keep a content calendar, stay visible. The advice is correct. The economics are not.
Hiring help is expensive: agency social media management runs €500-800+ per month, and even a freelance social media manager costs more than most micro-businesses can justify. Doing it yourself means evenings spent hunting for social media post ideas, writing captions, and slowly burning out. The graveyard of abandoned business accounts is enormous, and it isn’t a motivation problem. It’s a production problem.
Generic AI was supposed to fix this, and partially did. An AI caption generator or a chat assistant removes the blank page. But these tools are stateless: they don’t know the business’s offers, prices, clients, tone, or what happened last Tuesday. Ask a stateless model for a post and you get the statistical average of the internet, which is precisely why AI-written content has become so easy to spot. The output is fluent and belongs to nobody.
That gap, between “AI can write” and “AI knows what this specific business should say”, is a retrieval problem. Retrieval problems are our home turf.
The architecture: voice to social media posts, end to end
Laspi’s pipeline is short to describe and was not short to build:
- Once a week, the owner records a roughly two-minute voice note: what’s new, what arrived, what clients asked. Voice matters here, and not only for speed. Spoken language carries the owner’s real vocabulary, examples and rhythm, which typed prompts filter out.
- The note is transcribed and mined twice. It becomes this week’s raw material, and it quietly feeds durable facts into the business’s memory.
- A per-business memory layer (Tag-RAG over pgvector embeddings) retrieves what’s relevant: current offers, past cases, audience segments, objections, constraints.
- A marketing reasoning stage turns retrieved facts plus the weekly goal (“fill the quiet days”, “push the new service”) into a content plan, not a single post: an actual AI-generated content calendar for the week.
- Multi-format generation produces platform-specific captions for Instagram, Facebook, TikTok, LinkedIn, Telegram and others, images built from the owner’s real photos rather than synthetic stock, and short vertical videos with voiceover.
The owner reviews everything and publishes in a couple of taps. Laspi never asks for social media passwords; publishing stays in the owner’s hands by design.
One consequence of the pipeline shape: format follows the network, not the idea. A picture-first platform gets a square image with its own caption length and structure, a video-first platform gets a vertical reel, and nothing is copy-pasted across networks. “One post everywhere” is the most common small-business mistake, so we removed it architecturally.
The interesting part: business memory that compounds
The differentiator is not the generation. Generation is a commodity. The differentiator is that Laspi gets better the longer a business uses it, and that property comes from two RAG design decisions.
First, memory with a human checkpoint. The system proposes structured facts extracted from each weekly note (an offer changed, a new testimonial, a recurring client objection), and the owner confirms them. Only confirmed facts feed generation. Retrieval ranks by semantic similarity multiplied by a recency-decayed weight, so stale facts fade without being deleted: what the owner said this week counts in full, what they said a quarter ago counts a fraction. The memory stays current without maintenance.
Second, brand voice as a derived artifact, not a questionnaire. Nobody can fill in a form describing how they write. So Laspi infers voice attributes (formality, sentence length, signature phrases, words the owner would never use) from the owner’s actual posts and transcripts, refreshes that profile weekly from their own words, and enforces it as hard constraints during generation, with automatic drift detection when a draft slips out of character. This is the part that makes the output read like the owner on a good day instead of like an AI content tool.
Together these two loops answer the question every stateless assistant fails: not “write a post about coffee” but “write what this bakery, with these regulars, this new sourdough, and this way of talking, should post on Thursday”.
Trust constraints we treated as requirements
Three product decisions came directly from working with EU privacy and small-business realities:
- Real photos, deterministic text. Image models render only the visual scene from the owner’s own photos; headlines are burned on deterministically afterwards. No fake AI-stock look, no hallucinated storefronts.
- No credentials. The owner publishes themselves. An AI social media content generator for small businesses should reduce work, not collect passwords.
- Data locality by construction. Each business’s raw data serves that business’s content and is hosted in the EU. The architecture is designed so raw business data never leaves the individual business.
Where it stands
Laspi is in production with paying customers, generating content in English, Spanish and Russian, with the interface in the same three languages. Plans start at €19 per month, which reframes the usual arithmetic: instead of asking how much a social media manager costs, the question becomes whether a week of platform-native posts, images and video is worth less than a dinner out. The first week is free, without a card, and produces full-quality materials rather than a demo.
For Partenit, the project validated something we suspected from the dAIEDGE work: the value of retrieval-augmented systems is not in bigger models but in owning the context. A small model that knows the business beats a frontier model that’s meeting it for the first time, every single chat.
FAQ
What is an AI social media post generator with memory? A system that generates social media content from an accumulated, structured knowledge of one specific business (offers, clients, tone, history) instead of from a one-off prompt. Each generation retrieves relevant confirmed facts before writing, so posts stay accurate and on-voice.
How is this different from using ChatGPT for social media? A general chatbot is stateless: every conversation starts from zero, and output defaults to the average of its training data. Laspi maintains per-business memory and a derived voice profile, plans a full content calendar rather than a single caption, and produces platform-specific formats including images from real photos and short video.
How much does it cost compared to hiring a social media manager? Agencies typically charge €500-800+ per month and freelancers less but still substantially more than software. Laspi starts at €19 per month with a free first week.
Which platforms does it support? Captions and visuals are generated natively per network: Instagram, Facebook, TikTok, LinkedIn, Telegram, Threads and other picture- or video-first platforms, each with its own format and length.
Curious how it feels in practice? Record two minutes about your own business and generate a week of social media content from a voice note: the first week is free.
