Building an enterprise AI platform: permission-aware RAG that holds up in production
A reference architecture for enterprise AI: hybrid retrieval, user-scoped tools, fail-closed access control, event-driven deletes, and eval-gated CI/CD.
HLAB field notes
Blog
Analysis and field notes on the systems, tools, and ideas changing how modern software gets built.
55 published storiesPage 1 of 5
How I built a governed dependency pipeline: no package ships until it earns trust
A case study of a governed open-source intake pipeline: quarantine, Trivy scanning, OPA policy as code, promotion to an approved registry, and SBOM monitoring with Dependency-Track.
Score match quality and load results into Postgres
Use eval/score.py to compute precision, recall, and F1 for a reconciliation run against the ground_truth.json answer key, then optionally load the same run into a Postgres target schema.
Reconcile ERP and CRM records with the pipeline
Run the reconcile CLI to ingest, clean, validate, and resolve ERP and CRM exports into a single crosswalk, quarantine log, and Markdown reconciliation report.
Customize messiness and build a reproducible configuration file
Tune the missing-value, bad-date, encoding, duplicate, orphan, and amount-drift ratios that control how messy a generated dataset is, and save your choices to a config file datagen can replay exactly.
Compare ERP and CRM field mappings and the ground truth answer key
Read the same customer, invoice, and payment across a flat ERP CSV export and a nested CRM JSON export, and learn what the ground_truth.json answer key does, and doesn't, record.
Generate a synthetic ERP and CRM dataset
Use the datagen CLI to produce a reproducible, deliberately messy ERP and CRM export pair, complete with a ground-truth answer key, the foundation every reconciliation lesson in this series runs against.
Test model output with deterministic rules
Add deterministic rule checks to a prompt-evaluation workflow, then measure exactly where simple substring rules produce false positives and false negatives.
Benchmark retrieval quality
Turn six labeled retrieval questions into a repeatable BM25 quality report using Hit@k, Recall@k, and mean reciprocal rank, then use the report to find what lexical retrieval misses.
Merge rankings with Reciprocal Rank Fusion
Combine BM25 and vector rankings with Reciprocal Rank Fusion, promoting chunks found by multiple retrievers without ever comparing their incompatible raw scores.
Build and search a vector index
Build a small exact vector index that stores L2-normalized vectors, ranks results by cosine similarity, and rejects vectors from an incompatible model or dimension.
Implement Okapi BM25 from scratch
Build the keyword-ranking function behind a local retrieval pipeline: preserve identifiers, calculate Okapi BM25 scores, and return deterministic ranked results without a search library.