Armakuni Contact us
← All case studies

Many of its future members would rather speak
than fill in a form.

A Latin American neighborhood marketplace with more than 143,000 active users was about to launch in Mexico City, on AI that cost too much to scale and a sign-up form that shut out people less at ease with text. We moved its AI onto Amazon Bedrock, replaced the form with a conversation, then a spoken one in Latin American Spanish tested with the people it was for, and built an AI companion that remembers each member.

Before

The form was the barrier.

The AI bill grew with every neighbor.

After
  • Conversational registration, connected to the existing membership checks.
  • A voice interface in Latin American Spanish, tested with people from the audience it's for.

The form was the barrier

The marketplace connects neighbors to trade, find local services and talk in safer local networks, and it was growing across Latin America, from a test market in Argentina and production in Costa Rica toward a launch in Mexico City.

For a hyperlocal marketplace, the registration form was a business problem. Many prospective members could say what they needed easily enough but struggled with a text-heavy application. For some, speaking was the only practical way in, and the platform had no voice option.

The AI bill grew with every neighbor

The platform's AI already ran on third-party models outside AWS, and every conversation, moderation check and onboarding step drove cost through them, with no easy way to optimize. The Mexico City launch would multiply that demand. Moderation used basic rules with no sense of tone, so as the community grew, keeping it constructive got harder.

Let people talk, and let the platform remember

The plan moved the AI onto AWS first, with each task on the right-sized model, then let people join by talking, in text and then aloud, and finally gave each member an AI companion that remembers what they're looking for.

How we built it

5 decisions, in the order we made them. Each one says what we did and why, with what it did drawn beside it.

  1. Architecture checked against real users before anything is built
  2. Registration as a conversation that calls the real checks
  3. Voice tuned for Latin American Spanish, tested with the people it's for
  4. A companion that remembers
  5. Moderation that suggests a better way to say it

Architecture checked against real users before anything is built

Before anything is built Model usage1Volumes2Operations3Real users4 the architecture and cost model
The engineering

Before committing to a twelve-month, three-phase build, the business needed to know the architecture would work for its users, its markets and its roadmap:

  • The team inventoried the existing model usage, conversation volumes and marketplace operations, and mapped every AI workflow to its AWS equivalent.
  • Four to five discovery sessions tested the assumptions about how users behave, how many would use voice and whether conversational commerce was feasible.
  • The output was a target architecture for all three phases and a cost model tied to expected conversations and transactions, built on what discovery found rather than on list prices.

Why: before a city launch, the board needs costs based on how people actually used the service, and estimates don't give them that.

Registration as a conversation that calls the real checks

Built withBedrockAPI GatewayFargateLambda
A neighbour talksThe AI agent asksThe real checks runSigned up
The engineering
  • Bedrock-based dialogue collects identity information, location and community preferences, and keeps the context of the registration session.
  • Verification integrations connect the conversation to the existing checks needed for membership.
  • API Gateway is the entry point for every AI request, with rate limits and access controls, and Fargate services and Lambda handle the application workflows.
  • Models are chosen by how complex the task is, so simple interactions and harder reasoning use different capacity, and the migration included prompt optimization for each task.

Voice tuned for Latin American Spanish, tested with the people it's for

Built withTranscribeNova Sonic
Spoken SpanishTo text and backTested with the people it's for
The engineering

The voice experience extends the same onboarding workflow. Amazon Transcribe handles speech input, configured for the pronunciation and vocabulary of Latin American Spanish, and Amazon Nova Sonic provides natural spoken responses rather than robotic text-to-speech.

Testing used representative Latin American Spanish audio and people from the intended audience, including users with little experience of text-based interfaces. We wanted it to work for people who don't build software, and testing only with the team would have told us nothing about them.

A companion that remembers

Built withBedrockOpenSearchDynamoDBAPI GatewayLambdaFargate
The companion RemembersSuggests listingsSuggests events
The engineering

Bedrock embeddings and OpenSearch retrieve marketplace information, and the AI agent uses it alongside each member's explicit preferences and interaction history to suggest listings, community events and reminders. DynamoDB holds profile and application state, and retained conversation history lets a later session carry on where the last one stopped. Amazon API Gateway and AWS Lambda front the requests, AWS Fargate runs the MCP servers and the model engines, ElastiCache keeps hot data close, SQS carries the asynchronous orchestration and Secrets Manager holds the credentials, so the conversational path scales without a server to manage.

Moderation that suggests a better way to say it

A sharp postA gentler version offered
The engineering

Moderation runs as its own workflow: it assesses a post's sentiment, flags potentially problematic content and offers a constructive rephrasing for the member to consider, rather than just blocking it. Feedback informs later refinement of the moderation behavior.

What we'd tell another CTO

1Design for the member who can't use your form, because for some of this audience voice was the only way in, and test it with the real audience, since accent, vocabulary and confidence don't show up in a lab.
2Pick each model by its task, so simple dialogue doesn't pay for the biggest model, and price the AI on real behavior from discovery before you scale. In moderation, suggesting a calmer way to say something keeps both the member and the community, where a block keeps only one of them.

What else changed

  • The AI moved to Amazon Bedrock, across conversation, moderation and onboarding, with each task on a right-sized model and the prompts optimized.
  • An AI companion with memory, preferences and proactive suggestions.
  • Moderation that flags tone and suggests rephrasing.
  • An architecture and cost model planned for growth from 143,000+ active users toward one to two million, with CloudWatch monitoring and usage analytics across the new features.

What comes next

Conversational purchasing and transactions, such as an escrow-backed wallet for safe trades, so the companion can help complete a deal, not only suggest one.

How we'd do it today

This project came before our current offers. Here's how we'd approach the same work now.

At one to two million members most conversations are routine, so AI Inference Engineering would send them to a model the business owns on Amazon SageMaker AI, at about a tenth of the price, with a router sending each request to the smallest model that passes and the hard ones to a frontier model through Amazon Bedrock. The test would be in Latin American Spanish: a golden set of real conversations with the right outcome, which every change must pass, plus a sealed set nobody tunes to.

Before the companion could help anyone spend money, its AI agent would get its own identity in AgentCore Identity, with Cedar policies, spend limits included, checked outside the model on every tool call. Everything a member types would be treated as untrusted, because one hidden instruction in a message can steer an AI agent.

Modernize with AI AI Inference Engineering AK AI Factory

More work like this

  1. AI0A U.S. SMS marketing platform for restaurantsmanual CSV uploads, where campaigns used to wait on themIn SMS marketing, timing is the product. The data arrived when someone remembered to upload it.$0 net to the customer, for a $54,000 build AWS funded1 endpoint for every connected systemOn AWS Personalize · Lambda · S3 · CloudWatch
  2. AI40%A U.S. healthcare data platformfaster document processing, as reportedA misread dosage doesn't stay on the page. It follows the patient into the next system.35% lower extraction cost25% better text-extraction accuracyOn AWS ECS · Fargate · CloudWatch · Textract
  3. AIGreen, yellow, redA U.S. agricultural cooperativea verdict on the chemicals on screenThe mixing screen showed the chemicals. It never said whether they were safe together.4–5 weeks to a proof of concept that checks the chemicals on the mixing screen142 Safety Data Sheets in one searchable databaseOn AWS Nova Pro · S3 · Aurora · OpenSearch

All case studies →

Have a problem like this one?

A free workshop: a few hours with the architects who'd do the work, on your own systems. No cost.

Book a free workshop