Selected work

Public reference · Applied AI

Databricks LLM Content Pipeline with Human Review

A registry-driven workflow that screens source material, generates structured content, routes invalid output, and prepares valid results for human review.

Type
LLM production workflow
Role
Architect / engineer
Status
Public reference
Focus
Reliable generation and review operations
DatabricksPySparkDelta LakeLLM endpointsStructured generationRetriesUnity CatalogHuman review
System map

The operating path.

A deliberately simplified architecture view. The case study below explains where the important guarantees and decisions live.

  1. 01Content registry
  2. 02Source aggregation
  3. 03Deterministic screen
  4. 04Model fallback
  5. 05Structured generation
  6. 06Validate + review

The problem

Why this system needed to exist.

A model call is not a content system. Production generation also needs work selection, source assembly, relevance filtering, bounded context, retries, schema validation, state transitions, invalid-output handling, and a usable reviewer handoff.

Without those controls, teams spend their time chasing half-finished records and repairing outputs that look valid but do not satisfy the downstream contract.

The system

How the pieces work together.

A Delta-backed registry selects source-complete items and tracks generation state. Deterministic domain, keyword, title, and source heuristics remove obvious noise before model calls; ambiguous snippets can use a smaller relevance model.

Bounded context is sent to a larger generation endpoint. Required output structure is validated, valid and invalid files are routed separately, retries use exponential backoff, and optional Google Drive and Slack steps move approved files into a human-review workflow.

01

Deterministic before probabilistic

Rules handle the obvious decisions cheaply and consistently. Models are reserved for ambiguity and generation rather than used as a replacement for all control logic.

02

Invalid is a real state

Malformed output is routed and recorded instead of being silently accepted or disappearing inside a failed notebook run.

03

Human review is part of the architecture

The output contract includes stable files, reviewer delivery, status updates, and optional notifications - not just generated text in a model response.

Source boundary

The public version generalizes client names, workspace identifiers, prompts, schemas, source content, folders, and webhooks. Its local demo uses a deterministic fake model so the workflow can be inspected without Databricks or external inference.

A system like this on your roadmap?

Let’s make the hard parts explicit.

Discuss the project