Atlas Index

Context

For a platform distributing academic and cultural content, pure manual indexing no longer scaled. We designed an assist service: the system proposes normalized metadata from transcripts and domain vocabularies. Humans remain the only decision-makers before catalogue publish.

Challenge

Rising volume, uneven indexing, weak discoverability. Design goal: cut first-pass time without auto-publishing vocabulary errors (people, topics, rights).

Solution

Ingest (existing transcripts or ASR) → NER/topics/embeddings → controlled-list alignment → per-field scores → keyboard HITL queue → publish API only after validation → accept/reject logs to retune thresholds and prompts.

Baseline

Without assist: 100% manual indexing, time varies by item type, backlog growing faster than editorial capacity. The useful metric is not NLP latency alone; it is median minutes to a reviewable record.

Results

• Suggestions with per-field confidence scores
• No metadata published without human validation
• Reusable OpenAPI for CMS / player
• Learning loop from accepts and rejects (thresholds, not necessarily day-one fine-tune)

Architecture decisions

• Choice: Specialized extractors + LLM for normalization

  Why: Reproducibility and cost at volume; the LLM does not own the whole pipeline.

  Rejected: End-to-end LLM only (cost, non-determinism, weak vocabulary control)

• Choice: Alignment to controlled lists

  Why: Catalogue interop and editorial quality; free-text tags rot search.

• Choice: Mandatory HITL before publish

  Why: False positives on people/topics/rights = reputation risk and catalogue debt.

  Rejected: Auto-publish above a single global threshold

Limits

• Short-text NLP latency is not end-to-end time (ASR, HITL queue, review).
• Multi-language, diarization and slide OCR stay hard cases: isolate them in the backlog.
• Without a clean vocabulary, AI accelerates chaos: start with list inventory.

Checklist to rebuild this

1. Inventory vocabularies (people, topics, rights) and current record quality
2. Measure baseline: median minutes / item on 50-100 samples
3. Pick 1 content type for the POC (not the whole collection)
4. Define metadata schema and scored fields
5. Build the HITL queue before any fine-tune
6. Log accept/reject per field
7. Keep publish API separate from suggestion
8. Per-field thresholds + dual-review policy