A source-grounded publishing intelligence platform I designed to make fragmented international publishing data easier to research, verify, evaluate, and trust.
Independent portfolio/product project — not employer production work.
Publisher sites, catalogs, university presses, public announcements, language-specific sources, and rights-related records rarely use one consistent structure. A model can summarize this material fluently while still making unsafe or unsupported conclusions.
The product goal is to organize evidence without pretending that missing information proves a business fact.
I designed the product architecture, ingestion and review workflows, research-agent behavior, prompt governance, execution telemetry, AI evaluation, deterministic rights-safety controls, testing strategy, and technical documentation.
My quality-engineering background strongly shaped the architecture: every important AI behavior should be testable, observable, and reversible.
The model is one component. The surrounding platform provides data quality, provenance, safety boundaries, observability, and release controls.
Controlled tool use over structured WPH records and source evidence, with safe abstention when evidence is insufficient.
Immutable prompt versions, hashes, draft/candidate/production lifecycle, evaluation-gated promotion, audit history, and rollback.
Trace IDs, models, prompt versions, tokens, estimated cost, latency, failures, database-backed pricing, and workspace budgets.
Golden datasets, deterministic and model evaluators, regression comparison, failure taxonomy, human review, and blocking quality gates.
Deterministic enforcement prevents unsupported rights claims from being introduced by prompt or model changes.
Source discovery, parsing, normalization, staging, duplicate checks, proof-chain review, and controlled public promotion.
I deliberately separated the data pipeline from the AI execution layer. The model does not become the system of record.
A shared trace ID connects AI execution details across the request, prompt, model, evaluation, and telemetry layers.
Source discovery → fetch/parse → normalization → staging → duplicate/entity checks → review queue → proof-chain checks → verified promotion.
Primary Nordic development has covered Denmark, Sweden, Norway, Finland, and Iceland, including commercial publishers and university-press workflows.
External publishing data is inconsistent and can conflict across sources. WPH keeps discovered, staged, needs-review, verified, and public states separate so incomplete evidence does not silently become public truth.
Ingestion monitoring also tracks source failures, parser omissions, duplicate rates, zero-staging anomalies, and source health.
I built evaluation into prompt promotion and regression testing rather than treating it as a dashboard added afterward.
Safe: “No English edition was found in the reviewed sources.”
Unsafe inference: “English rights are available.”
Those statements are not equivalent. Rights safety is a hard gate and cannot be averaged away by strong scores elsewhere.
Model scores and human scores remain separate. Evaluation results link back to prompt versions, model execution, trace IDs, cost, latency, and failure categories so a regression can be investigated rather than merely observed.
backend tests passing at a completed stabilization milestone
backend subtests passing at the same milestone
frontend tests passing, with lint, TypeScript, and production build also green
These are milestone-specific verified counts, not a claim that the continuously evolving repository will always contain the same number of tests.
I kept one primary system of record rather than adding specialized databases before the product required them.
Safety-critical rules stay outside model judgment so an evaluator model cannot “vote away” a prohibited claim.
The current RAG work starts with immutable source snapshots, deterministic chunks, entity provenance, canonical URLs, hashes, and trust tiers before adding vector retrieval.
These are the four WPH demos currently in the portfolio. Contextual embeds elsewhere on the site point back to this same set; they are not additional videos.
Best starting point for technical recruiters and hiring managers. A product-first walkthrough of the implemented WPH platform.
Watch product demo ↗For engineering managers and senior engineers: system boundaries, prompt governance, telemetry, evaluation, testing, and tradeoffs.
Watch architecture deep dive ↗For AI evaluation, quality, and applied-AI teams: deterministic checks, model evals, regression testing, human review, and hard safety gates.
Watch evaluation deep dive ↗A broader project overview with career context, architecture, research-agent workflow, observability, and lessons learned.
Watch recruiter briefing ↗