CASE STUDY 02 · MARKETPLACE
Nexus ODM
An AI-assisted review tool that checks long documents against compliance rules and points reviewers to the exact clauses that need attention.
ROLE
Full-stack + AI engineering
STACK
NestJS · BullMQ · pgvector · Claude
TIMELINE
[x weeks]
TYPE
Client project
01 — PROBLEM
Compliance reviewers read long documents line by line against a rulebook that keeps changing. Reviews were slow, findings were inconsistent between reviewers, and there was no record of why a clause was flagged.
02 — ARCHITECTURE
Upload UI
documents + rulesets
NestJS API
auth · jobs · results
BullMQ
review queue
Workers
parse · chunk
PostgreSQL + pgvector
chunks · embeddings · rules
Claude API
clause-level review
Findings report
clause · rule · citation
03 — KEY DECISIONS & TRADE-OFFS
Queue every review instead of answering in-requestDocuments can be hundreds of pages. BullMQ workers process chunks in parallel, retry failures and report progress to the UI.TRADE-OFF More moving parts: Redis, worker scaling and job-state UI.
pgvector inside Postgres, not a separate vector DBRules, documents, embeddings and findings live in one database with one backup and transactional writes.TRADE-OFF Less tuning headroom than a dedicated vector store at very large scale.
Claude reviews clause by clause, with citationsSmall, retrieved context per clause keeps answers grounded and lets reviewers see exactly which rule triggered a flag.TRADE-OFF More API calls per document; mitigated with caching and batching.
04 — STACK
NestJS
API + job orchestration
BullMQ
Review queue + workers
pgvector
Rule ↔ clause retrieval
Claude API
Clause-level review
05 — RESULTS
[metric]
Review time per document, before vs after.
[metric]
Share of findings accepted by reviewers.
[metric]
Documents processed per week.
NEXT CASE STUDYFilipina Dating App →