Case study · Operator cockpit

Ana Concierge OS

A local-first cockpit for case work, in which an assistant can draft, cite and propose, and a person approves every action.

Role
Sole developer
Built with
TypeScript, React, Fastify, PostgreSQL, zod, Vitest
Status
Working local foundation; outbound actions deliberately out of scope
Layers
Interfaces, systems, models, data and evaluation
Fig. 1 · One assistant turn, from question to audit trail
  • Built by me
  • A person, or an external model
  • OperatorA person
  • Ana serverBuilt by me
  • ModelVia a bounded runner
  1. Operator to Ana server: Asks about a case
  2. Ana server: Rebuilds context from PostgreSQL
  3. Ana server: Retrieves passages within scope
  4. Ana server to model: Prompt and passages; no shell, no general tools
  5. Model: Drafts an answer
  6. Model to Ana server: Schema-valid JSON: answer, citations, proposed action
  7. Ana server: Rejects unretrieved citations and out-of-scope actions
  8. Ana server: Hashes the proposal and queues it
  9. Ana server to operator: Answer with sources; proposal awaiting approval
  10. Operator: Approves or rejects
  11. Operator to Ana server: Decision
  12. Ana server: Records it in the audit trail
Context

The boundary belongs in the architecture

Concierge-style work is a stream of cases: people, requests, follow-ups, documents, and the knowledge needed to answer well. An assistant can help with all of it, provided it never invents a source and never acts on its own.

Ana Concierge OS is my attempt to build that boundary into the architecture rather than into a prompt. It runs locally, keeps its state in PostgreSQL, and treats the model as a drafting tool whose output the server checks and a person approves.

What I built

Six decisions that keep the assistant honest

  1. 01 · Systems

    Cases with rules and a record

    Cases, relations, tasks and follow-ups live in PostgreSQL behind a typed Fastify API, with deterministic case rules, daily work queues and an audit trail.

    I built the Fastify API on Node.js, with zod validation of route parameters, query strings and request bodies. The /v1/* routes require a bearer token.

  2. 02 · Systems

    Context rebuilt on the server

    Every assistant turn starts from canonical state rebuilt from the database, not from whatever the conversation happened to remember.

  3. 03 · Data

    Retrieval with a scope

    Knowledge sources are normalized, chunked and reviewed before they are activated. Search uses PostgreSQL full-text, filtered by workspace, case, review state, date and classification, and every retrieval leaves a trace.

  4. 04 · Models

    A model on a short leash

    The local runner starts the model client without a shell, passes the prompt on standard input, disables general tools, requires schema-valid JSON and stops on timeout.

  5. 05 · Systems

    Checked citations, human approval

    The server rejects a citation to any passage that was not retrieved, and any action outside the conversation’s scope. Proposals are hashed and wait in an approval queue for a person to decide.

  6. 06 · Data

    Imports that stay local

    Document imports are local, bounded, verified by hash and staged until reviewed, with no model call involved.

Interface

A cockpit that shows its uncertainty

A responsive React interface with a today view of priorities, plus cases, relations, knowledge, imports, approvals and a persistent assistant. The assistant’s answers show their citations, uncertainty and suggested actions instead of hiding them.

A synthetic discovery mode lets someone explore the whole product without any operational data.

Verification

Tests where the boundary is

Unit and integration tests cover approvals, assistant safety, knowledge retrieval, document-import privacy and acceptance, and isolation of the model runner. Smoke scripts exercise the API and the assistant end to end.

Status and scope

What it is, and isn’t

Status. A working local foundation for case work; I don’t quote usage or adoption figures for it.

Scope. Outbound connectors for messaging, email and calendars, and any autonomous action, are deliberately outside it. Retrieval is lexical PostgreSQL search; embeddings may come later.