AI answers you can defend in court.

Regulated industries don't lose because their AI is wrong. They lose because nobody can prove it was right. Winnex returns every answer with a signed record showing the search was exhaustive — and that nothing was skipped.

Book a walkthrough Read the benchmarks
100%Recall@10
0Bound violations
7.1×Faster than HNSW
LGPDEU AI Act · GDPR

The problem

An AI that can't show its work is a liability, not an asset.

Every vector database on the market returns an answer and asks you to trust it. Approximate search skips documents. Nobody knows which ones. When a regulator, a judge, or an auditor asks “did you actually look at everything?”, the honest answer today is: nobody can prove it.

We built the answer to that question.

The platform

Three layers. One guarantee.

Winnex is not a wrapper around someone else's model. It is the retrieval engine, the audit record, and the products built on top of both.

Deterministic search

A C++20 engine that examines every document in the corpus and returns a mathematically bounded result set. No approximation, no skipped rows, no silent recall loss.

Engine architecture

Proof of audit

Each query produces a signed, tamper-evident record: which corpus, which threshold, what was excluded and why. Exportable as a certificate that survives legal review and WORM storage.

Audit architecture

Products that ship

Tracer-GOV, Tracer-JUS, Tracer-MED and Tracer-GAP put the engine to work in government, legal, clinical and financial triage — each with its own domain model and its own evidentiary chain.

See the Tracer Series

The Tracer Series

Mathematical-proof retrieval for regulated domains.

Four products, one engine. Each query returns the answer and a signed record of what was examined and what was excluded — exportable as a certificate an auditor can check without us.

Preprint and benchmark results are published on Zenodo (DOI 10.5281/zenodo.21826158, latest version v4) with the raw results in the open.

Components from $24,000/yr per component, small tier
Full RAG Proof $306,000/yr normalize + madhava + tracer
Tracer-GOV Free Brazilian public sector
See full licensing

Tracer-GOV

Public-sector audit. WORM-backed procurement and policy review that survives freedom-of-information scrutiny. Frameworks: LGPD, EU AI Act.

Benchmark on Zenodo

Tracer-JUS

Legal e-discovery. Every document examined, every exclusion justified, delivered with a certificate in PDF for the case file.

Benchmark on Zenodo

Tracer-MED

Clinical triage. Evidence-backed triage with a record of what the model saw and what it ruled out. Framework: HIPAA, LGPD.

Benchmark on Zenodo

Reading the benchmarks honestly

  • Bound pairs equal corpus size × queries — this is a linear scan with mathematical bounds, not a sublinear index. The point is correctness and auditability, not speed.
  • Recall@10 = 1.0000 against exhaustive search is expected by construction: nothing is skipped, so nothing can be missed.
  • WORM storage is tamper-evident, not tamper-proof.
  • The engine is licensed BSL 1.1 — source-available, not OSI open-source.
  • Exclusions are proven outside the top-K by vector similarity, not by relevance judgement.
  • float32 arithmetic requires an ε-slack in the bound; float64 needs none.
Compliance and regulatory review

Why it matters now

The compliance wave is already here.

The EU AI Act makes documentation and traceability a legal requirement for high-risk systems. LGPD and GDPR require you to explain how a decision was reached. “The model returned it” stopped being an answer.

  • Traceability by construction. The record is written at query time — not reconstructed later, when it is already too late.
  • Runs where the data lives. On-premise, air-gapped or in your own cloud. Nothing has to leave the jurisdiction.
  • No vendor lock on the answer. The proof is a signed artefact you own, verifiable without us.
100%Recall@10Measured against exhaustive search, not against a margin.
0Bound violationsAcross every benchmark suite in the corpus.
7.1×Build speed50K documents: 0.086 s against 0.61 s for HNSW.
13Benchmark tablesPublished, reproducible, with raw results in the open.

Where it runs

Built for the places where being wrong is expensive.

The same engine powers four products. What changes is the corpus, the review rules, and the shape of the certificate each domain needs.

Clinician reviewing patient records

Tracer-MED — clinical triage

Evidence-backed triage with a record of what the model saw and what it ruled out.

Public sector compliance review

Tracer-GOV — public sector audit

WORM-backed procurement and policy review that survives freedom-of-information scrutiny.

Legal and document review

Tracer-JUS — legal e-discovery

Every document examined, every exclusion justified, delivered with a certificate in PDF.

Financial reconciliation and gap analysis

Tracer-GAP — financial gap analysis

Reconciliation across ledgers where a missed line is a real loss, not a rounding error.

“The question is no longer whether the model is accurate. It is whether you can prove what it looked at.”
The premise behind every Winnex product
Winnex AI operations

The company

Born inside an ERP. Tested on real workloads.

Madhava, our search engine, did not start as a research project. It was built to be the retrieval layer of Winnex Maestro, our own business management platform — which means it has been answering real commercial queries under load since day one.

Winnex Brasil Soluções Empresariais LTDA-ME is a Brazilian deep-tech company. Our engineering is published, our benchmarks are reproducible, and our models run on infrastructure you control.

GitHub Documentation

Get started

Bring provable certainty to your AI stack.

We run a technical walkthrough on your own corpus — your documents, your questions. You leave with the recall numbers and a signed certificate, not a slide deck.

Prefer emailcontato@winnex.ai → Evaluate firstTechnical docs →

We reply to the email you provide. No newsletter, no CRM sequence.