A/IAALQUIM IASELECTED WORK · CASE 02
← Back to portfolio

CASE 02 · TECHNICAL PILOT DOCUMENT INTELLIGENCE

AuditorIA

From a scattered regulatory corpus to answers with retrievable context.

AuditorIA organizes technical documents, creates semantic representations and retrieves relevant fragments before composing an answer. The goal is not to replace expert judgment, but to make the corpus more queryable and traceable.

ROLEARCHITECTURE · RAG · PRODUCT2026
Open AuditorIA View service health
2,767items in the observed index
16standards reported by the service
BGE-M3embedding model
RAGretrieval pattern

PROJECT CARD / RECRUITER VIEW

The project's technical dossier.

EXPLICIT STATE TECHNICAL PILOT
TARGET USER

Consultants, auditors and technical teams that need to query a regulatory corpus without losing the link to the source.

WHAT I BUILT

I designed the RAG architecture, the separation between retrieval and generation, the query surface and the pilot's evidence criteria.

AI ENGINEERING SIGNAL / RECRUITER VIEW

Backend, APIs, deployment, testing and controls.

DEPLOYMENT
Public application inside the Alquim-IA AI Suite, served over HTTPS.
BACKEND / API
API under /oracle/ai-suite/api with public health and regulatory search.
AI STACK
Index of 2,767 items, BAAI/bge-m3 embeddings and a deepseek-v4-pro LLM reported by the service.
TESTING
Health 200, 'residuos' search, catalog and history verified; end-to-end PDF analysis still pending.
OBSERVABILITY
Visible API state and an explicit separation between retrieval, generation and reporting.
ARCHITECTURE OVERVIEW
  1. 01Documents + metadata
  2. 02BGE-M3 embeddings
  3. 03Semantic index
  4. 04Retriever
  5. 05LLM + references
WHAT WORKS TODAY
  • Public interface
  • Health endpoint
  • Index with reported content
  • Assisted queries over the corpus
HARD TECHNICAL DECISIONS
  • Separate health, retrieval and substantive correctness
  • Keep the expert in the loop
  • Never confuse a fluent answer with sufficient evidence
MEDIA AND DEMO
DEMO
Public application
SCREENSHOTS
The live interface works as visual evidence
VIDEO
Pending recording
WHAT IS STILL PENDING
  • It is a technical pilot: service health alone does not prove the correctness of every answer.
  • The observed corpus and its coverage can change; the figures correspond to the stated snapshot.
  • Regulatory decisions still require professional review and verification against the primary source.

01 / PROBLEM

The information exists, but finding it is not the same as understanding it.

Technical teams work with standards, studies and annexes that vary in form, length and vocabulary. Keyword search leaves context out; answering without traceability introduces a bigger risk.

02 / DESIGN DECISION

A retrieval layer before generation.

The system separates ingestion, normalization, embeddings, search and composition. Each stage can be observed and evaluated independently instead of hiding everything behind a single prompt.

SYSTEM / END TO END

The product as an explicit flow.

Each phase has a distinct responsibility. That separation makes it possible to observe, test and fix the system without relying on intuition.

  1. 01

    Ingest

    Convert heterogeneous sources into processable units with identity and metadata.

    INGEST
  2. 02

    Represent

    Create multilingual embeddings to approximate meaning, not just literal matches.

    EMBED
  3. 03

    Retrieve

    Select relevant passages within a controlled context window.

    RETRIEVE
  4. 04

    Compose

    Generate an answer grounded in the retrieved material and preserve its references.

    ANSWER
  5. 05

    Evaluate

    Separate service health, retrieval quality and substantive correctness of the answer.

    VERIFY

PROOF / WHAT IS ACTUALLY SHOWN

Evidence, not decoration.

The labels indicate whether the data was observed, comes from the artifact or belongs to the professional record of the case.

OBSERVED · AUG 11 2026HTTP 200public interface accessible
OBSERVED · AUG 11 2026STATUS OKhealth endpoint
SERVICE RESPONSE2,767 / 16items and standards reported
OBSERVED CONFIGURATIONDeepSeekLLM reported by health

SCOPE / HONEST BOUNDARIES

What is demonstrated and what still needs validation.

A technical portfolio must also show where the available evidence ends.

  • It is a technical pilot: service health alone does not prove the correctness of every answer.
  • The observed corpus and its coverage can change; the figures correspond to the stated snapshot.
  • Regulatory decisions still require professional review and verification against the primary source.

CASE 02 / AUDITORIA

From a scattered regulatory corpus to answers with retrievable context.

Open AuditorIA View the full portfolio