Giovanni Beijl
← All work

Cortex

AI infrastructure built so an answer stays attached to its evidence, and can refuse to answer at all.

2025–present built solo in production

Overview

Cortex is a multi-tenant AI platform for organizational knowledge: retrieval-augmented generation over a customer’s own documents, with citations on every answer and mandatory human review before anything reaches a customer.

Context

Support and knowledge teams want AI answers drawn from their own documentation, but an answer that sounds confident and is wrong is worse than no answer. The system needed a way to tell the difference between “I found this in your documents” and “I am guessing,” and to treat those two cases completely differently.

Constraints

No accuracy benchmark or golden dataset exists yet for this system. That is a genuine current limitation, not a caveat to bury: the architecture is built to fail closed by design, so that the absence of a benchmark is a gap in measurement, not a gap in the actual answer-or-refuse decision.

Role

Designed and built solo, as a standalone product under The Canopy Corporation: platform architecture, retrieval, ingestion pipelines, the verification path, and the API surface.

Architecture

How Cortex decides whether to answerA question triggers hybrid retrieval across semantic, keyword and fuzzy search over the customer’s own documents. Retrieved evidence is checked before generation; if the check fails, the system returns an explicit no-answer. If it passes, an answer is generated and then a separate verifier pass can still fail it closed. A passing answer is returned with citations to the source passages, and a human reviews the draft before it reaches a customer.questionsemantickeywordfuzzyhybrid retrieval · customer documents onlyevidence checkgeneratepassverifier passanswer + citationspasshuman reviewto customerno answer, explicitlyno supporting evidencefails closed

Plate IHow Cortex decides whether to answer at all, and the two separate places it is allowed to decide not to.

Decisions

  • Verification is a separate pass from generation, not a prompt asking the same model to grade its own work.
  • Human review is mandatory before any answer reaches a customer, not an optional setting.
  • Refusal is an explicit, first-class output, not an error state.

Alternatives

Deliberately did not ship model-generated answers without citations, even though it would have been faster to build and would have looked more capable in a demo.

Implementation

  • Ingestion pipelines for JSONL, PDFs, codebases, and documentation libraries, with embedding backfills and reindexing
  • An OpenAI-compatible API with streaming responses over Server-Sent Events
  • Model training workflows for QLoRA, LoRA, and full fine-tuning, including adapter-aware inference

Reliability

No golden dataset or retrieval benchmark exists yet. That is a real limitation of the current system, stated plainly rather than implied away with confident language.

Results

In production for B2B support teams on Freshdesk, with every customer-facing reply passing through the evidence check, the verifier, and a human before it ships.

Lessons

The harder engineering problem was never retrieval. It was building a system honest enough to say no answer, and building it so that saying so is the default outcome of a missing check, not something a developer has to remember to add later.