Every answer names the note it came from. No note, no answer.
MIA Human Technology builds vertical AI assistants for small and mid-sized firms in Brazil's regulated sectors: law, accounting, healthcare and hospitality. Retrieval is one layer of ten. The other nine are what keep the answer honest.
What this company builds
- Company
- MIA Human Technology, a software company based in Brazil. Founder: Daniel Ravetta.
- Product
- Vertical AI assistants for law, accounting, healthcare and hospitality firms.
- Method
- Ten layers around retrieval: a curated note injected into context, vector and semantic search, expansion of the matched chunk to its mother note, a similarity graph, a rerank layer, and a base that cleans and rewrites itself. Gaps are filled by notes generated from official sources, reviewed in public by the Legion community, then approved in batches by the MIA team. Each answer names the note it used.
- Limit
- When no curated note covers the question, the agent abstains instead of answering anyway.
- Control
- Sensitive output waits for human sign-off. Every access is logged. What travels is pseudonymized and only the user's key reads it back. A local mode puts base and inference on the client's own hardware: tested at volume, not yet in beta.
- Stage
- Products run in beta, testing or development. None of them is in general production.
Retrieval is one layer. These are the other nine.
Vector search and semantic search run inside this system. That part is not the argument. Naive retrieval fails in ways any team that has shipped it recognizes, and each failure has a layer against it. The unit underneath all of them is a curated note: anchored to a verifiable source, tagged by vertical, and admitted to the base only after a human has reviewed it.
The path, in words: a curated note is injected into the context, vector and semantic search run, the matched chunk is expanded to the mother note it came from, a similarity graph links notes to one another, and a rerank layer separates the notes that answer from the notes that only scored. The answer names the note it used. When nothing reliable comes back, the agent abstains, and abstention becomes the second trigger on the same auto-generation layer that already reacts when an official source changes: one mechanism, two ways in. What it writes is reviewed in public by the Legion community, approved in batches by the MIA team, and only then does it enter the base.
-
The chunk reaches the prompt without the context it came from.Mother-note expansion. When a fragment matches, the system climbs to the note that fragment belongs to and reads that note whole.
-
The corpus rots. A note goes stale and nothing flags it.Auto-cleanup and auto-generation. Obsolete content is removed on its own. When an official source publishes a change, the note for it is generated from that source and reviewed by the Legion community before it counts.
-
Ranking by similarity returns what scores well, not what answers.Context engineering rerank. A layer above retrieval separates the notes that actually answer the question from the notes that merely scored.
-
To be useful, sensitive data has to leave the building.A base on the client's own machine. The local mode keeps the base and the inference on the client's hardware, so what cannot leave does not have to. What does travel goes up pseudonymized, readable only with the user's own key.
- WikiMIA inject. The curated note, source-anchored and human-reviewed, enters the context.
- Retrieval. Vector search and semantic search run over the note base.
- Mother note. The matched chunk is expanded to the note it came from.
- Similarity graph. Notes are linked to one another, not held as loose fragments.
- Rerank. A context engineering layer separates what answers from what only scored.
- Auto-cleanup. Obsolete content is removed from the base automatically.
- Auto-generation. When an official source changes, the note is generated from it and reviewed before it counts.
- Own brain. A brain built per client, from the data only that company has. tested, not yet beta
- Local base and local inference. Both can sit on the client's own hardware, for data that cannot leave it. tested, not yet beta
- Pseudonymized upload. What travels can only be read back by the user, with a cryptographic key they hold.
When no curated note covers the question, the agent says it does not know.
Ten layers can still come back with nothing trustworthy. When that happens, an answer with no note behind it is not produced.
The gap then closes in three moves. The missing note is generated from an official source. It is reviewed in public by the Legion community, which anyone can join: people who want to work on this, including lawyers who help other users check a contract an AI wrote, and lawyers already running AI inside their own practice. Then the MIA team approves the reviews in batches, with the MIA copilot.
An open door and a closed gate, and the review in the middle happens in the open, where it can be audited. The next answer has a source, nobody had to sit down and write the note by hand, and no unreviewed text ever counts as knowledge.
The firms that use these assistants work under professional secrecy and under LGPD, Brazil's data protection law. An assistant that invents a citation or exposes a document is a liability. Abstention is the cheaper outcome, and it is a designed one.
Agent graphs, not a single prompt.
Agent graphs run on LangGraph, with tracing. Each product composes several agents with defined roles, and each step of the run is recorded.
In MIA Resort, for example, Mario handles sales and service, Maia handles rate and management from real data, Maisa handles marketing, and Sam holds the security and approval layer.
Every access is logged. Sensitive output waits for a human.
SAM is the governance layer embedded in every MIA product. It is not sold separately.
It shows the source of each answer, logs every access with user, time and purpose, and requires human sign-off before a sensitive action leaves the system. It works as a circuit breaker: a confident answer does not become a document or a course of action without a person saying yes.
The data that cannot leave the desktop does not leave it.
Professional secrecy and LGPD are the constraints these firms work under. What holds up in front of a lawyer is architecture, not adjectives. Here is the architecture.
On the client's machine
The local mode puts the base and the inference on the client's own hardware, so data that cannot leave does not have to. Tested at volume on rented GPU infrastructure. Viable and validated, not yet in beta.
Pseudonymized, and only the user holds the key
Data is pseudonymized before it reaches the model, and whatever goes up travels pseudonymized. Reading it back takes a cryptographic key, and that key is the user's. Not a promise that we will not look: an arrangement where we are not the ones who can.
Credential encryption
Credentials are encrypted. We state the scope precisely: this is credential encryption, and it is in operation.
Row Level Security
What a client puts in their own knowledge base is isolated by RLS. Only that client reaches the database of their own documents.
Infrastructure we do not share
Postgres and Supabase run self-hosted on a virtual private server of our own, not inside a third party's multi-tenant service. The processors are dedicated, with no other tenant on them.
Human sign-off
Sensitive output does not leave the system without human approval, and every access is logged with user, time and purpose.
Models are a dependency, not the product.
Each role is bound to an inference API behind an adapter, so a model can be swapped — or an entire provider repointed — without touching the note base or the audit trail. That portability is deliberate: the notes are the asset, and they outlive any model generation.
The system does not need an external API to work.
Tested at volume on rented GPU infrastructure · not yet in betaFrontier models over an API
Each role is bound to a frontier model through its API. This is how the products run today, including the ones listed as beta.
Open models in the 30B class
The same system runs on open models in the 30B class, with no external API call. In our own testing, quality on these tasks holds against the frontier APIs. That is a result from our bench, not a published benchmark.
Where the local mode actually stands. We tested it at volume on rented GPU infrastructure. That proves the open-model path holds under load. It does not prove an installation running on a client's own desktop, and we are not going to say it does. Viable and validated; we will call it beta when a client is running it, and not before.
Along the way we ran several models collaborating on a single graphics card, with good results. Multi-agent orchestration is not a diagram here. It fits on modest hardware.
This is what the architecture is built for. For a firm under professional secrecy, the sentence stops being "your data leaves protected" and becomes "your data does not have to leave".
Next on the bench, and not built yet: a planned experiment integrating models through a shared latent vector. It is listed because it is planned, not because it runs.
What runs today, and at which stage.
Stages as recorded on 28 July 2026. None of these products is in general production, and we do not describe them as such.
MIA Resort
Hotels and restaurants: rate, direct booking, marketing, and food and beverage.
miaresort.wikimia.cloud →MIA Law
Law and accounting firms. Issues Brazilian tax filings (DAS, PGDAS-D, DCTFWeb) through the SERPRO Integra Contador integration, with an A1 digital certificate. Knowledge indexed by area of law.
mialaw.wikimia.cloud →MIA Legion
Platform for building and running sites, stores and web apps.
legion.wikimia.cloud →MIA Finance
Accounting firms: tax routine and client communication.
miafinance.wikimia.cloud →MIA Med
Clinics and medical practices: scheduling and patient communication.
miamed.wikimia.cloud →MIA Forge
An engine that runs coding loops: one agent writes the prompt, another implements, a checker verifies, and the cycle repeats under a governance gate. Human approval moves up one level at a time: report, then assisted, then autonomous.
MIA Realestate
Vertical for real estate brokers.
MIA Commerce
Vertical in design. Scope not published yet.
SAM
Security, trace and approval. Embedded in every product above, not sold separately.
We publish the stage of each product because a claim of production that an evaluator can disprove costs more than the claim is worth.
Questions engineers ask first.
Is this RAG, or something else?
Retrieval is one of the layers: vector search and semantic search run here. The difference is what surrounds them. A curated note is injected first, a matched chunk is expanded to the mother note it came from, a similarity graph links notes to one another, a context engineering layer reranks what actually answers, obsolete content is cleaned automatically, and when an official source changes the note for it is generated and reviewed before it counts. The claim is not that retrieval is absent. The claim is that retrieval alone is not enough.
What happens when there is no reliable source?
The agent abstains and says it does not know. The missing note is then generated from an official source, reviewed in public by the Legion community, which anyone can join, and approved in batches by the MIA team with the MIA copilot. Only after that does it enter the base. The next answer has a source, nobody wrote the note by hand, and nothing unreviewed counts as knowledge.
Does a human review sensitive output?
Yes. SAM, the governance layer embedded in every product, requires human sign-off before a sensitive action leaves the system. It also shows the source of each answer and logs every access with user, time and purpose.
Where does client data live?
Today, on Postgres and Supabase, self-hosted on a virtual private server and CPU we do not share, with credentials encrypted and Row Level Security isolating each client's database. What travels is pseudonymized, and reading it back takes a cryptographic key that the user holds. A local mode, with the base and the inference on the client's own hardware, is tested at volume but not yet in beta.
Can this run without sending anything to an external API?
Architecturally yes, and it has been tested, but we do not call it delivered. What runs today is the default mode, where each role is bound to a frontier model over an API. The local mode runs the same system on open models in the 30B class with no external API call; we tested it at volume on rented GPU infrastructure, and in our own testing quality on these tasks held against the frontier APIs. That is a result from our bench, not a published benchmark, and it does not prove an installation running on a client's own desktop. Viable and validated, not yet in beta. We have also run several models collaborating on a single graphics card. Integrating models through a shared latent vector is a planned experiment, not something we run today.
Which models does the platform use?
Four roles: reasoning, multimodal, embeddings and video. Each role is bound to an inference API behind an adapter, so a model can be swapped — or an entire provider repointed — without touching the note base or the audit trail.
What is the current stage of each product?
MIA Resort, MIA Law and MIA Legion are in beta. MIA Finance, MIA Med and MIA Forge are in testing. MIA Realestate and MIA Commerce are in development. None of them is in general production. Stages as recorded on 28 July 2026.
Talk to us.
Tell us the sector you work in and which question your team keeps asking. We will walk you through the beta closest to it, show where each answer comes from, and where the system abstains.
WhatsApp is the channel we answer on, in Portuguese or English.