Technology
Small models, tuned to one market.
A retrieval and rewriting system built for continuous work on one country’s e-commerce data. Designed on AWS. Fine-tuning already runs on NVIDIA H100 GPUs; embedding, reranking, training and serving at volume move onto NVIDIA NeMo Retriever, NeMo and NIM next.
Built on AWS · fine-tuning on NVIDIA H100 today · NeMo Retriever, NeMo and NIM next
Five stages, one loop.
Signals come in from the web, the store and AI assistants. Rewrites go out to the storefront. Every outcome comes back as data for the next pass.
Crawl + events
Scheduled crawls and store events become normalised signals per product.
Embed + rerank
Pages are embedded and re-ranked for recency and market relevance.
Market-tuned small LMs
Fields rewritten from retrieved evidence and the product’s own data.
Versions + rollback
Connectors write the new version to the storefront and keep the last one.
Outcomes per version
Results are logged per version and become preference data for the next tune.
05 Learn feeds 02 Index and 03 Rewrite on the next pass
What each stage does.
CPU where the work is plumbing. GPU where the work is understanding and writing at volume.
Crawlers read search results, competitor listings and reviews on a schedule; store events (stock, price, new products) arrive as they happen. Everything is normalised into signals tied to one product and one city.
Thousands of pages per category are embedded and re-ranked, so the rewrite step sees the evidence that matters for this product this week, not last month’s.
The model writes the new title, keywords, description, answer copy and structured data, in each enabled language. A check rejects any claim the product data does not support.
Connectors write the version to the storefront through its API and keep the previous version for one-click rollback. The first rewrite per store waits for the owner’s approval.
Impressions, clicks, add-to-cart, sales and assistant citations are logged against each version. Kept and rolled-back versions become preference pairs for tuning.
The model plan.
Start with hosted models on Amazon Bedrock to learn what a good rewrite looks like. Then move the rewrite path to small open-weight models tuned for Nigerian listings, trained and served on NVIDIA GPUs.
Why GPUs.
The work is continuous, not one-off: the market changes daily, and every product is rewritten on every signal.
Continuous embedding
Thousands of competitor and marketplace pages per category are re-embedded and re-ranked on a schedule, not once.
Per-market fine-tuning
Small open-weight models tuned on Nigerian product copy in five languages, retuned from logged outcomes. The plan is training, not prompting.
Rewriting at scale
Every SKU, on every signal, several fields, several variants. Batched GPU inference keeps a rewrite cheap enough to run continuously.
Data handling.
Stores hand Akoni their catalogue and their results. The system is designed so that trust is easy to check.
One store, one boundary
Each store’s catalogue, analytics and logs are kept separate. One store’s private data never trains a model for another store.
On AWS, encrypted
Store data will be kept in the AWS region selected for the product, encrypted at rest and in transit.
Leave with your data
A store can ask for its data and its version history, or ask for both to be deleted. See the privacy notice.
Where it stands.
What exists today, and what comes next. Dates are plans.
Fine-tuning on H100
Open-weight models fine-tuned on NVIDIA H100 GPUs. Built on AWS. Early access next.
Training on NeMo
Fine-tuning moves onto NVIDIA NeMo; NeMo Retriever for the index.
Tuned models in the loop
NIM serving for the rewrite path; results published per language.
Items marked planned are roadmap. Platform and product names belong to their owners and imply no partnership or endorsement.