Engineering evidence · 26 October 2026
Inspectable code and recorded checks: a restricted hosted warehouse, an AI guide baseline that keeps the failures, and optional project recommendation signals.
01 / Data engineering
Upstash remains the operational store. A batch exporter reads daily aggregates into PostgreSQL tables and SQL reporting views. Browser identifiers and chat conversations stay out of the warehouse.
Existing recorded traffic, isolated by project and environment.
Transactions, replay protection, corrections and coverage checks.
Missing data stays unknown. Partial days stay visible.
Synthetic demonstration · not customer traffic
35 recorded views across two available closed days in this fixture. One day is unavailable; the five partial-day views are excluded. Daily unique browsers are never added into weekly unique people.
Verification boundary
Fourteen analytics tests cover replay, corrections, rollback, isolation and restricted permissions. Live Supabase TLS, row reconciliation and replay checks passed.
The first manually dispatched production import passed. Daily scheduling is enabled; the first clock-triggered execution remains unobserved in this report. Real counts stay private and belong to the separate no-bloom project.
Remaining: backup restoration, load testing and external stale-import alerts. These checks do not establish customer impact.
Inspect hosted verification ↗02 / Applied AI
The live portfolio guide retrieves approved public sources before requesting an AI answer. A baseline run tested factual questions, unsupported requests, injection attempts and Vietnamese questions.
One pass, no retries. Nine citation-bearing responses used approved URLs. Median latency was 1.137 seconds; p95 was 2.684 seconds, including failed requests.
The 16 questions came from a 24-case authored development set. URL and phrase checks are regression signals, not a factual-accuracy score. Token usage and cost were unavailable.
When generation fails after validation and rate checks, the guide can show a clearly labelled source excerpt with a citation. Failure tests cover that behavior; rate-store outages still fail closed.
The published baseline predates this fix. A complete post-release live evaluation is pending. The report preserves scoring limitations and all four failed requests.
03 / Recommendation engineering
Visitors can choose to share project-link interactions from the guide’s “Help improve project recommendations” settings. Collection starts off and can be stopped at any time.
Recent aggregate counts can break close relevance ties. Explicit project questions keep their original retrieval order. Sparse, stale or unavailable signals leave ordinary retrieval unchanged.
The language model receives approved evidence, never private counts. This is aggregate ranking, not a profile of an individual visitor.
Tests cover consent, expiry, allowlisted destinations, storage failures and ranking boundaries. An isolated Redis check verifies concurrent deduplication and rate limits without creating production engagement.
No recommendation-quality uplift is claimed. Redis interaction signals are separate from the PostgreSQL traffic warehouse. Neither vector search nor model training is used.