Sell with Marketing replaced a static AI file upload with a live, reference-anchored knowledge system on Claude for ESCG, a professional services firm. Every answer now comes back anchored to a numbered source inside a confidential corpus that keeps growing — not to whatever happened to be sitting in context that day.

The problem: a corpus too big, and too alive, for a file upload

ESCG works with a large, confidential document corpus, held under formal reference numbering, cryptographic hashing, and timestamping, and it keeps expanding as new material comes in. That combination — size, confidentiality, and continuous growth — is exactly what breaks a normal AI setup.

The first approach was to load the files directly into a Claude project. It ran into three structural limits at once. Project knowledge has a capacity cap, and a corpus this size didn't fit. It's static, so it couldn't grow with new material. And because everything sat in context at the same time, the model answered from whatever was loaded instead of the verified record — returning answers without the source references the firm needs to stand behind its work.

What we built

SWM came in as Claude's implementation partner. We audited the existing setup, rebuilt the knowledge base with Claude Code in VS Code, and replaced the static uploads with a private, live retrieval layer. It exposes three tools to the firm's Claude project: corpus search with excerpts anchored to their reference numbers, plus persistent write and read access to a working memory, stored in a private repository. The corpus never touches a public surface, and the firm keeps full ownership of the repository.

The Claude stack

Three Claude products, working together:

  • Claude Code (the VS Code extension), for building and distilling the corpus.
  • Claude.ai Projects, as the end-use surface the firm's team works in every day.
  • A custom MCP server, built on the Anthropic SDK (Node/Express, deployed on Render), connecting the Claude project to a private, version-controlled corpus on GitHub.

What changed, in numbers

Results, verified against the client's own repository:

  • Corpus integrity verified: 158 of 158 primary files checked against a SHA-256 manifest before ingestion.
  • 254 documents and 80 portal pages distilled by Claude Code into a structured 22-file, 9.55 MB knowledge base, split into an analytical layer and a 14-package thematic corpus.
  • 12 standardized output templates generated from the distilled corpus.
  • Capacity ceiling removed: the static upload became a live retrieval layer, so the corpus grows without re-uploading files or manual reindexing.
  • Every answer returns reference-anchored excerpts, so the user checks each fact against its numbered source instead of trusting the model's memory.
  • Persistent memory writes decisions back to the repository, giving continuity across sessions and across team members.
  • An optimization pass cut response time substantially: one recursive Tree API call instead of per-file requests, parallel batch reads of eight, a shared in-memory cache with TTL and write invalidation, and a warm-up endpoint that removes cold starts on the host.
  • Fully private architecture: no confidential material leaves the client's own GitHub and hosting accounts.

Why this matters beyond one firm

Any professional services firm sitting on a confidential, formally referenced corpus that keeps growing runs into the same three walls: a size cap, a static snapshot, and answers with no way to check the source. The fix isn't a bigger upload — it's a retrieval layer that treats the corpus as a living, versioned system instead of a one-time attachment. The same discipline that makes an answer verifiable to a person is what makes it citable at all: a reference a human can check is also a reference a model can point to.

Verified, on the record

This case is published with the client's direct authorization. Ramón Escovar at ESCG confirmed the details above and can be reached at rescovar@escg.com.ve to verify them.

This is what we build: AI systems on Claude for process improvement, data analysis, and putting a company's own knowledge to work. Get in touch and we'll tell you plainly whether it fits your case.