Machine learning · 2026

InterfaceGenie

Retrieval-grounded document and deck generation.

Generated documents fail in a specific way: they are fluent and unsourced. The platform is built around the opposite constraint — retrieval first, generation second, with the corpus as the authority and the model as the writer. Three product modes share the pipeline: a corporate reporting wedge, a lighter student and creator mode, and a professional planning mode for designers.

Status
Active
Domain
Applied LLM · Document systems
Role
Architecture, RAG pipeline, product
Year
2026

Stack

  • RAG
  • Model Context Protocol
  • Vector search
  • Postgres
  • Stripe
  • TypeScript

A platform that generates corporate reports and presentation decks from a governed corpus, with a retrieval pipeline, tool integration through the Model Context Protocol, and a planning mode for designers.

01

Retrieval as the authority

The generation stage is deliberately constrained.

Content is retrieved and ranked from a governed corpus before any prose exists, and the model composes within that evidence rather than around it. The architecture pack that documents this treats the retrieval pipeline, the data model and the cost analysis as first-class design artifacts — because in a generation product the retrieval quality and the unit economics are the product.

02

Tooling through an open protocol

External capability is wired through the Model Context Protocol rather than bespoke integrations, so a new data source becomes a server rather than a fork of the application.

It is a slower first integration and a dramatically cheaper tenth.

03

Staged from local to cloud

The build is explicitly sequenced: a fully local stage that proves the pipeline on a single machine, then a cloud stage that adds multi-tenancy, billing and scale.

The order means the expensive infrastructure decisions are made against a working system rather than a guess.

04

Three modes, one pipeline, different failure costs

The corporate wedge, the creator mode and the professional planning mode share retrieval and generation but not their tolerance for error.

A student deck that misattributes a statistic is embarrassing; a corporate report that does the same is a liability. The modes therefore differ in how much of the pipeline is allowed to be automatic — the corporate path keeps a human verification surface over every retrieved claim, and the lighter modes trade that for speed knowingly rather than by omission.

05

Cost per artifact is a design input

A generation product where each document costs an unbounded amount of inference does not have a business model, it has a bill.

Token accounting is treated as an architectural constraint from the start: retrieval depth, context assembly and model selection are all tuned against a target cost per artifact, and the analysis lives beside the architecture documents rather than being discovered at scale.

06

Security as a documented workstream

Payment handling, secret management and history hygiene are tracked as explicit artifacts alongside the feature work rather than as a pre-launch checklist.

For a product taking payment and holding client documents, treating the security review as a deliverable — with its own specification, its own findings and its own remediation record — is the difference between a claim and a position you can defend.

Pipeline

Retrieval before generation

Stage Can reject

  1. 01

    Corpus

    Governed sources

  2. 02

    Chunk + embed

    Vector index

  3. 03

    Retrieve

    Ranked evidence

  4. 04

    Compose

    Draft

    Model writes within the evidence

  5. 05

    Bind citations

    Sourced claims

  6. 06

    Render

    Document or deck

Skills exercised

What the build
actually demanded.

Against the corpus · 15 systems

  • Stack breadth 6
  • Design decisions 6
  • Pipeline stages 6
  • Decision gates 1

This system Corpus median

Retrieval

  • Chunking and embedding strategy for long-form source documents
  • Hybrid ranking over a governed corpus
  • Citation binding between generated claims and sources
  • Evaluation of retrieval quality independent of generation

Product

  • Multi-mode product scoping from one pipeline
  • Subscription billing and entitlement design
  • Designer-facing planning and production workflows
  • Cost modelling per generated artifact

Platform

  • Model Context Protocol server integration
  • Staged local-then-cloud build sequencing
  • Relational data model for documents and revisions
  • Security review and secret-handling remediation

What it establishes

  • Three product modes

    Corporate · creator · professional planning

  • Protocol-native tooling

    New sources are servers, not forks

  • Local stage first

    Infrastructure chosen against a working pipeline

  • Cost per artifact budgeted

    Token accounting is an architectural constraint

  • Security as a deliverable

    Specified, reviewed and remediated on the record