Interview Unified Portal

Interview scores are a single 1–5 rating; review uses 7 categories × 1–10 — the unified ranking averages both

Fix: pick one scale for both stages before interview day — recommendation: 7 categories, 1–5, matching existing form copy — so ranking numbers become comparable.

Reviewers see all 123 applicants — no assignment scoping, no progress tracking

Fix: per-reviewer applicant assignment table + filter + progress bar. Personnel note (Sept 14): the evaluation committee and interview panels may be different people, and interviewer rosters can differ per session date — reviewer assignments and interviewer assignments are TWO separate models (reviewer→applicants; interviewer→rooms per session as the schema already supports). The single biggest reviewer-experience upgrade for review week.

Zero tests on scoring math, assignment generation, rotation timers, date filters

Fix: add vitest; start with the extracted date-helper (pairs with P0-3 fix) and score-aggregation tests.

1.58 MB single JS bundle; no route-level lazy loading

Fix: lazy() per domain (admin / interviewer / kiosk / evaluation) — same pattern that cut an Admin chunk on Unified Platform from 450 KB to ~80 KB.

No backup automation for season data; rollback is a manual dist.bak

Fix: nightly pg_dump of urology_* to a VPS volume. 2 hr.

Interviewer day-of flow doesn't surface paper-stage context by default

Fix: render 3-line paper-stage highlights on the room card — makes "one system" feel real to faculty in the room.

Cycle archive captures counts, not data exports

Fix: archive step writes a CSV/JSON export of the closing cycle.

No 48h interviewer prep email pipeline

Fix: cron-driven email from that session's own interviewer roster (rosters can differ between dates) + assigned applicants + 3-line paper highlights — prefaced "from the evaluation committee" since interviewers may not be reviewers. 4–6 hr.

Sees all 123 applicants, self-tracks progress

/evaluation → raw list, no assignment scoping, no progress indicator

Needed this week: assigned-only view + progress bar (P1-2) — review week is 5–10 reviewers × ~12 applicants.

Best surface in the app — timed rotations, realtime updates

dashboard → schedule → room → score

Paper-stage context exists in ApplicantProfileDialog but sits one click outside the default flow (P1-7) — surface it on the room card.

Tooling-rich, but cross-stage work feels like six separate tools

6 admin pages + interview mgmt + ranking, no progress ribbon

Admin ribbon (Review % → Interview Day → Ranking %) is the design fix.

Nothing — no portal, no itinerary

planned P3

Applicant itinerary portal (arrival instructions, room rotation, day timeline) is on the 6-month roadmap.

Shape

Single SPA → Supabase direct (no BFF). Auth via UnifiedAuthProvider + has_urology_role() RPC; three roles (admin/interviewer/evaluation). Cycle context (InterviewCycleContext) defaults to hardcoded '2025-2026'.

Component inventory

Boundary problems

1. InterviewSystem.tsx = 3542L god component — evaluation UI, review API calls, categories, DNR logic all in one file.

2. No service/repository layer — 26 components call supabase.from('urology_*') inline (268 total call sites), no central query builders.

3. Hardcoded session-date logic (28 sites) in 5 files — business rule "Einstein-only on Oct 30/31; special flow Nov 24/Dec 11" is buried in components against literal calendar dates. Breaks the moment interview dates change (i.e. this season).

4. Legacy login pages (3 login flows + unified) still in bundle.

Lifecycle data flow (verified in code)

urology_reviews (paper stage, reviewer_name TEXT) → read in ApplicantProfileDialog + ConsolidatedRankingBoard + InterviewerRoomView (paper context before interview) → urology_interview_scores (single rating 1–5 per room, notes, question_comments jsonb) → ConsolidatedRankingBoard merges avg review + avg interview per applicant → urology_final_selections (rank_order, decision). Chain EXISTS but interview scale ≠ review scale (see 32).

Method: source sync from VPS (312 files), grep sweeps, build/tsc probes in throwaway copy.

Findings

After-lens verdicts

High-risk untested logic: assignment generation+overwrite, rotation timers/pause math, score aggregation (rounded avg formulas), Einstein date filters, RLS-protected admin flows.

No REST backend: all data ops = Supabase client calls direct.

Edge functions (10)

import-applicants, import-new-applicants, clean-and-import-applicants, add-single-applicant, upload-applicant-pdf, upload-single-pdf, pdf-proxy, bootstrap-admin, create-interviewer-auth, reset-interviewer-password

Client query surface (verified counts)

Cross-ref issues

Radix/shadcn primitives + a genuinely strong kiosk visual system (glow-success, kiosk-stat-card-compact, animate-scale-in, dark/light kiosk themes); tailwind token usage mostly present; large-type day UI reads well at distance.

Weak points: raw slate-* classes bypass tokens in interview components; AdminReports (939L) density risks clutter; the zero-PDF empty state is not designed (immediately relevant given materials gap).

Design strategy — this season

① ONE applicant card across all 3 stages (review → interview → ranking): same card, stage-scoped sections. Faculty feel one product. ② Kill raw slate-* → consolidate tokens. ③ Admin progress ribbon: Review % → Interview Day → Ranking %, persistent across admin pages. ④ Designed empty states for the zero-materials review queue ("Waiting on materials — 123 applicants pending") instead of blank cards. ⑤ Kiosk stays photo-free (anon SELECT surface — keep it that way when photos arrive).

Score: 4.0

Score: 2.0 (P1)

Score: 6.0 (no live Lighthouse — negotiate cap 8.0; mechanical evidence supports 6)

Auto-import note (Sept 14)

Extraction pipeline is LOCAL-ONLY: PyMuPDF/pdfminer + regex + tesseract OCR, all verified installed on the VPS — no AI provider, no external API, sensitive applicant data never leaves the server.

Key journeys traced in code + bundle (no live-browser walkthrough this run — PARTIAL):

1. Reviewer review: login → /evaluation → pick applicant → 7-cat score + notes → save. Friction: no assignment-scoping → reviewer sees all 123.

2. Interviewer day: login → interviewer dashboard → schedule → room view → timed rotation → 1–5 rating + notes → submit. Timer + reset working. Paper-stage context only if admin opens profile dialog — not in the default interviewer flow (opportunity).

3. Admin ranking: /consolidated-ranking → drag-rank (dnd-kit) → final_selections write. avg review + avg interview MERGED but scale mismatch makes averages lie (CR-08).

4. No applicant-facing portal (planned gap).

Dead ends: legacy login routes (/admin-login etc) redirect to /login — good.

Every finding the Aug 7-8 ce:full recorded is listed below with TODAY'S status (verified by the Sept 14 probes). Nothing from the accepted report is dropped: resolved items are kept as proven fixes, open items stay live findings.

P0 — Critical (6, August)

P1 — High Impact (7, August)

P2 / P3 (from August — still-relevant items retained)

August Roadmap — as written then, fulfilled (for the record)

- Week 1 (Stabilize & Secure): D1 compose · D2 create historical cycle row · D2 cycle filtering · D3 RLS harden · D4-5 remove departed reviewer + clean auth users + remove service-role key → all executed Aug 7-8 except compose (still open)

- Week 2 (Integrate & Launch): D6 yearly setup wizard · D7 CSV import tool · D8-9 merge Review into /evaluation · D9 wire kiosk/interviewer to real data · D10 cycle selector → executed Aug 7-8 (merge verified; wizard + cycle rows exist)

August 6-Month Vision (quarters as accepted then)

Q4 2025 (as planned): Stabilize & Integrate — DONE. Q1 2026: Unified portal + cycle mgmt — DONE. Q2 2026: Real-time interview day + consolidated ranking + analytics — built, data-empty (this season will exercise it). Q3 2026: Archive cycle · analytics dashboard · reviewer email pipeline — email pipeline still open (P1-9). Q3–Q4 2026: automation + year-2 wizard — superseded by Sept 14 season plan (P0-6 auto-import etc.).

August Feature Upgrades + ROI Ranking (status Sept 14)

August Feature Upgrades (the forward-looking list, then → now)

Edge-case matrix: whom does the date filter include/exclude?

Rating/averaging math (verified)

Policy: what happens when…

Lifecycle integration (the "one system" goal, in code)

Missing: per-reviewer reviewer-assignment completeness report (who hasn't submitted what) — the one admin report needed daily during review week. Per-session day-of summary with quotes (competing feedback for the PD) — flagged as design-opportunity, not bug.

User stories

Season Readiness (13 SP): admin — current cycle 2026-2027 (0.5) · reviewer — PDFs visible inline (5) · dev — date rules config-driven (2) · admin — stray test data purged (0.5) · ops — containers in compose (5).
Unify Scale (8 SP): one scale across 7-cat review and day-of interview so ranking compares like with like.
Reliability & Tests (13 SP): vitest on date-helper, score aggregation, assignment overwrite; CI basics.

Realtime publication excluded every interview table — subscriptions ran but never received events

All 13 live components (room monitor, kiosk, interviewer room view, leaderboard, scoring progress) subscribe via postgres_changes. The subscriptions opened fine but got no events — the "live" dashboard looked dead because, downstream of the socket layer, it was.

Fix (applied to the live DB Sept 14): added the 5 interview tables to the publication (rotations, scores, assignments, rooms, sessions) — verified in pg_publication_tables; realtime container restarted healthy. This is a configuration fix, minutes not days.

Validation: before interview day, run one end-to-end socket test (insert a test rotation, observe a subscribed client react). That closes the loop with proof instead of hope.

If realtime misbehaves on the day: 5-second polling is already built in

RoomMonitor already has a 5s polling backup alongside its subscription (RoomMonitor.tsx:243). The built plan: extend that poll-fallback to the other day-of surfaces behind a toggle — live = realtime when healthy, degraded = 5s polling when not. Either way the room monitor and kiosk reflect the room.

Recommendation: don't skip the live dashboard. The root cause was configuration and is now fixed; the safest posture is realtime primary + polling fallback, validated with a live socket test before the first session.

Evaluation committee and interview panels are separate people — and interviewer rosters can differ between session dates

Two separate assignment models: reviewer→applicants (paper stage) and interviewer→rooms (per session, via the existing room interviewer slots). No global season roster is assumed anywhere.

Interviewer context: for an interviewer who wasn't on the review committee, the paper-stage highlights on their room card are prefaced "from the evaluation committee" — it's their first exposure to the applicant's paper record, not a reminder.

Per-session rosters: the 48h prep email and session views pull that session's own interviewer list. A person can interview on one date and not the other — the schema (per-room interviewer slots) already supports this; the tooling just can't assume continuity.

Reporting stays role-split: reviewer completeness (who hasn't finished paper reviews) and interviewer coverage (who's assigned to which rooms/rotations) are reported separately — a session-only interviewer having zero review rows is correct, not a red flag.

Login was completely broken in production — auth pointed at a dead Supabase Cloud host

bundle contained fsxzgkyiutpjobpmfgtk.supabase.co (DNS NXDOMAIN) while the real backend is supabase.srv1738752.hstgr.cloud — "Failed to fetch" on every sign-in attempt

Three hardcoded cloud URLs lived in src (AdminSetupInterviewers.tsx ×2, PDFViewer.tsx). Fixed Sept 14: repointed to the self-hosted stack, rebuilt clean, container restarted, verified login → /admin/dashboard. The portal could not be logged into at all before this audit.

Method: authenticated headless Chromium walkthrough of all 16 routes (login → route loop, screenshots + computed-style harvest), page-by-page vision review, animation counts from getComputedStyle. Complements source-level 10a/10b — this is RENDERED evidence.

P0 discovered during the walkthrough (fixed same day)

Login was completely broken in production. The deployed bundle pointed auth at fsxzgkyiutpjobpmfgtk.supabase.co (an unresolvable Supabase Cloud host — DNS NXDOMAIN), while the self-hosted stack is supabase.srv1738752.hstgr.cloud. Symptom: "Failed to fetch" on every Sign In. Three hardcoded cloud URLs lived in src (AdminSetupInterviewers.tsx ×2, PDFViewer.tsx). Fixed: sed to self-hosted host, rebuilt (clean), restarted container, verified login → /admin/dashboard. This is the single highest-impact finding of the whole audit: the portal could not be logged into at all before today.

Font system (rendered)

Verdict: font choices are coherent and modern; NOT a font problem. The problem is rendered SIZE and weight on kiosk labels (10–12px — unreadable at distance).

Motion / animation inventory (computed, not claimed)

The app is effectively animation-free in rendered output. Kiosk NEEDS motion (pulse, counter roll, bar fill) because "live" must be visible.

Per-page verdicts (rendered)

Cross-cutting design themes

1. Admin surfaces are clean-but-generic (Bootstrap-era card+pill vocabulary). Acceptable for internal use; not a differentiator. The quickest premium lift: density toggle, sticky headers, search/filters, semantic color on KPIs.

2. Kiosk is the one screen with a real design concept (dark display) — and it's the least finished. Fixing kiosk (scale ×2.5, live motion, empty states, dedup tabs) is the highest-value design deliverable for interview day.

3. Empty states are the systemic gap. Rooms, interviewers, comments, ranking, kiosk all fail the "is the system alive or broken?" test. A shared EmptyState component (icon + reason + one CTA) fixes six screens.

4. Data-trust UI issues undermine otherwise-good screens: the 118-vs-123 KPI mismatch and the ranking empty-state copy are the two places a faculty user loses confidence in 5 seconds.

5. Motion: the design system has animation classes in source, but rendered output is static. Either wire them (kiosk first) or stop shipping them.

Design sprint recommendations (interview-season priorities)