Discover
Define the workflow, trusted notes, and out-of-scope decisions.
INDEPENDENT FORWARD-DEPLOYED ENGINEERING CONCEPT · 2026
ClientOps
with guardrails.
A proof-of-work slice for Palete’s AI Engineer role: turn approved design-project notes into cited answers, measurable checks, and a clear human handoff.
TypeScriptNext.jsPythonREST APIRAG · baseline
An independent concept, not Palete client work
The workflow and project records are fictional. It is not affiliated with Palete. The live demo uses a deterministic retrieval baseline; it does not call an LLM or claim production client outcomes.
LIVE PROOF OF WORK · FICTIONAL DESIGN ENGAGEMENT
Client delivery desk
A small, bounded workflow for answering project questions from approved notes.
P
PAL / PROJECT DESKApproved notes only · no action-taking tools
EXTRACTIVETry a sample question
Response
Ask a question to see the retrieval, evidence, and safety decision.
!
Human control stays in the loop
NO API KEYPermit, safety, contract, price, and delivery commitments are routed to a person. Unsupported questions get no invented answer.
No model key or external AI service is used. This demo is a deterministic retrieval baseline, not a generative AI assistant.
01 / SCOPE
A bounded delivery workflow
The prototype answers routine questions about a fictional design engagement. It returns a cited excerpt, declines unsupported questions, and escalates requests that could create a commitment.
Retrieve
Validate a small question, rank approved notes, and return source evidence.
Evaluate
Run golden cases for source selection, abstention, and escalation.
Hand off
Keep financial, legal, safety, permit, and schedule commitments with a person.
02 / ARCHITECTURE
The smallest useful architecture
A Next.js route handler provides a same-origin JSON API, validates bounded input, and returns a no-store response. A deterministic retrieval module keeps the behavior repeatable without model credentials. A Python standard-library runner calls the same golden-set endpoint used by the workbench.
03 / EVALUATION
How quality is checked
Six golden cases exercise the important path: retrieve the expected note, cite evidence, abstain when the corpus has no answer, and route risky requests to human review. The result is visible in the demo and executable from Python against local or deployed environments.
04 / OPERATIONS
Operator handoff
The operator receives the source excerpt and an explicit decision state. The assistant cannot send messages or change a CRM. Scope changes, prices, contract language, permits, safety, and promises about dates remain with a delivery lead.
05 / LIMITATIONS
The model boundary is explicit
This build intentionally has no generative model or provider key. It demonstrates retrieval, citations, guardrails, and regression evaluation; it is not evidence of a production LLM rollout. A real client deployment would add tenant permissions, source ACLs, prompt-injection defenses, model evaluations, audit retention, and a human approval step before external actions.