Architecture deep-dives for 4 more projects — source-verified, adversarially checked
Fanned out the architecture deep-dive surface (piloted on talkak) to Mimi, DocVault, Beside, and ScreenBridge. Each was extracted from the actual local repo by an analyze → adversarial-verify → write pipeline; the verify pass cut or corrected a real wrong number on every project before it shipped. Every claim cites real files; every number is measured. EN now; the loader falls back to EN for KO until KO versions are written. ScreenBridge and Meta-Smart-Glass: ScreenBridge done; Meta hit a rate limit mid-run and is the one remaining.
What I did
Each project's /projects/<slug>/architecture page now carries a source-verified deep-dive in the same shape as the talkak pilot: a measured-metrics table, a request-path mermaid diagram, the core mechanisms with the real file names inline, and an honest "what this doesn't do" section.
- Mimi — the async LLM pipeline pinned off the DB connection (
@TransactionalEventListener(AFTER_COMMIT)+@Async, short-tx bookends around the network call,ObjectProviderself-proxy), a boundedCallerRunsexecutor, the Gemini→OpenAI→Claude fallback with a hand-rolled retry, repository-layer multi-tenant isolation, and token-version JWT revocation. - DocVault — the audit hash chain enforced in Postgres triggers (with a
pg_advisory_xact_lockto close a read-then-write race), chunked authenticated AES-GCM with a nonce layout that defeats reorder and truncation, two-tier ingestion auth, and a server-enforced operator assistant. - Beside — one codebase that scales local→prod by env var, dual cookie/bearer auth, from-scratch Apple JWT verification, owner-gated presigned media with HTTP Range.
- ScreenBridge — the AX/OCR-over-LLM coordinate fusion, the privacy chokepoint, and the irreversible-action gate.
How it stayed honest
Each project ran through analyze → adversarial verify → write. The verify stage re-opened every cited file and re-ran every measurement, and it caught a real error on every single project:
- ScreenBridge: the image downscale cap was claimed at 1568 px from a stale code comment; the live constant is 1024. Deny-list 19→17; irreversible keywords 40→35; the "five-layer privacy" framing softened because two layers are still in-progress in the code.
- Mimi: a multi-tenant finder count of 39 was a substring/call-site count, corrected to 25 declared signatures; an unverified "Spring Boot 3.3" version pin was cut.
- Beside: a "native push not implemented" limitation was removed because the current mobile code (
expo-notifications,getExpoPushTokenAsync, a scheduler Expo branch) refutes it. - DocVault: a LOC metric labeled "non-test" actually counted tests; corrected.
These are exactly the numbers an engineer reading the repo in an interview would have caught. Better the adversarial pass caught them first.
What's left
- Meta Smart Glass — done (commit
9f85201). The first run's verify was rate-limited; a workflow resume finished it (verify clean, 10/10 metrics reproduced, BLE-bridge-not-built called out honestly). All 6 projects now carry a deep-dive. - KO — the deep-dives are EN; the architecture loader falls back to EN for KO readers, so nothing is broken, but native KO versions are a follow-up.
Verified by
npm run build clean (all /projects/[slug]/architecture routes prerender, no MDX crash from the generics/diagrams); on a fresh local server all five architecture pages return 200 and the project pages now show the "Architecture →" link.
Commit
69e91ab