Instead of 10 sequential GPT calls (which timeout with GPT-5.1 on 99
projects), use a two-phase approach:
Phase 1 - AI Scoring: ONE API call asks GPT to score each juror's
affinity for all projects, returning a compact preference matrix with
expertise match scores and reasoning.
Phase 2 - Algorithm: Uses AI scores as the preference input to a
balanced assignment algorithm that assigns N reviewers per project,
enforcing even workload distribution, respecting per-juror caps, and
filling coverage gaps.
Benefits:
- Single API call eliminates timeout issues
- AI provides expertise-aware scoring, algorithm ensures balance
- Truncated response handling (JSON repair) for resilience
- Falls back to tag-based algorithm if AI fails
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Cap maxTokens at 12000 (was unlimited dynamic calc that could exceed model limits)
- Replace massive EXISTING array with compact CURRENT_JUROR_LOAD counts and
ALREADY_ASSIGNED per-project map (keeps prompt small across batches)
- Add coverage gap-filler: algorithmically fills projects below required reviews
- Show error state inline on page when AI fails (red banner with message)
- Add server-side logging for debugging assignment flow
- Reduce batch size to 10 projects
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Problems:
- GPT only generated 1 reviewer per project despite N being required
- maxTokens (4000) too small for N×projects assignment objects
- No fallback when GPT under-assigned
Fixes:
- System prompt now explicitly explains multiple reviewers per project
with concrete example showing 3 different juror_ids per project
- User prompt includes REVIEWS_PER_PROJECT, EXPECTED_OUTPUT_SIZE
- maxTokens dynamically calculated: expectedAssignments × 200 + 500
- Reduced batch size from 15 to 10 (fewer projects per GPT call)
- Added fillCoverageGaps() post-processor: algorithmically assigns
least-loaded jurors to any project below required coverage
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Root cause: batches of 15 projects were processed independently -
GPT didn't see assignments from previous batches, so expert jurors
got assigned 18-22 projects while others got 4-5.
Fixes:
- Track cumulative assignments across batches (feed to each batch)
- Calculate ideal target per juror and communicate to GPT
- Add post-processing rebalancer that enforces hard caps and
redistributes excess assignments to least-loaded jurors
- Calculate sensible default max cap when not configured
- Reweight prompt: workload balance 50%, expertise 35%, diversity 15%
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Fix project documents not displaying on jury project page (rewrote MultiWindowDocViewer to use file.listByProject)
- Add working download/preview for project files via presigned URLs
- Display project tags on jury project detail page
- Add autosave for evaluation drafts (debounced 3s + save on unmount/beforeunload)
- Support mixed criterion types: numeric scores, yes/no booleans, text responses, section headers
- Replace inline criteria editor with rich EvaluationFormBuilder on admin round page
- Remove COI dialog from evaluation page
- Update AI summary service to handle boolean/text criteria (yes/no counts, text synthesis)
- Update EvaluationSummaryCard to show boolean criteria bars and text responses
- Add evaluation detail sheet on admin project page (click juror row to view full scores + feedback)
- Add Recent Evaluations dashboard widget showing latest jury reviews
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Round open/close scheduling now runs as a 60s setInterval inside the
app process (via instrumentation.ts register hook) instead of needing
an external crontab. Removed the /api/cron/round-scheduler endpoint.
- DRAFT rounds auto-activate when windowOpenAt arrives
- ACTIVE rounds auto-close when windowCloseAt passes
- Uses existing activateRound/closeRound from round-engine
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Auth layout verifies user exists in DB before redirecting to dashboard,
breaking infinite loop for deleted accounts with stale sessions
- Jury/Mentor layouts handle null user (deleted) by redirecting to login
- Filtering stats cards and result list now use effective outcome
(finalOutcome ?? outcome) instead of raw AI outcome
- Award eligibility evaluation includes admin-overridden PASSED projects
- Award shortlist reasoning column shows full text instead of truncating
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Projects with no files are now automatically rejected before AI
screening runs, regardless of whether a DOCUMENT_CHECK rule is
configured. This prevents the AI from incorrectly passing projects
that have no supporting documents.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The Re-evaluate button was producing fewer eligible results because
the standalone job sent minimal project data (title, description, tags)
while the integrated filtering pass sent full data (files, team, institution).
- Fetch rich project data in award-eligibility-job (files, team, institution, etc.)
- Relax AI prompt to be inclusive like the integrated pass — strong primary
criterion fit is sufficient, don't require all dimensions above 0.5
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Single AI call now evaluates both screening criteria AND award eligibility.
Awards with useAiEligibility + criteriaText are appended to the system prompt,
AI returns award_matches per project, results auto-populate AwardEligibility
and auto-shortlist top-N. Re-running filtering clears and re-evaluates awards.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add Rounds tab to award detail page with create/list/delete functionality
- Add "Entry point" badge on first award round (confirmShortlist routes here)
- Fix round detail back-link to navigate to parent award when specialAwardId set
- Filter award rounds out of competition round list
- Add specialAwardId to competition getById round select
- Warn on confirmShortlist when no award rounds exist (SEPARATE_POOL mode)
- Remove auto-tag rules from award config, edit page, router, and AI service
- Fix competitionId not passed when creating awards from competition context
- Add AUTO_FILTER quality threshold to AI filtering dashboard
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add missing fields to FilteringConfigSchema (aiParseFiles, startupAdvanceCount,
conceptAdvanceCount, notifyOnEntry, notifyOnAdvance) — Zod was silently
stripping them on save
- Restore auto-save with 800ms debounce on config changes
- Add staggered animations for filtering results (stream in one-by-one)
- Improve AI screening prompt: file type label mappings, soft cap handling,
missing documents = fail, better user prompt structure
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Part 1 - Bug Fixes:
- Fix toProjectWithRelations() stripping file fields needed by AI (detectedLang, textContent, etc.)
- Fix parseAIData() reading flat when aiScreeningJson is nested under rule ID
- Fix getAIConfidenceScore() with same nesting issue (always returned 0)
Part 2 - Special Award Track Integration:
- Add shortlistSize to SpecialAward, qualityScore/shortlisted/confirmed fields to AwardEligibility
- Add specialAwardId to Round for award-owned rounds
- Update AI eligibility service to return qualityScore (0-100) for ranking
- Update eligibility job with filteringRoundId scoping and auto-shortlist top N
- Add 8 new specialAward router procedures (listForRound, runEligibilityForRound,
listShortlist, toggleShortlisted, confirmShortlist, listRounds, createRound, deleteRound)
- Create award-shortlist.tsx component with ranked table, shortlist checkboxes, confirm dialog
- Add "Special Award Tracks" section to filtering dashboard
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add aiPreview mutation with full project/juror data (bios, descriptions,
documents, categories, ocean issues, countries, team sizes)
- Increase AI description limit from 300 to 2000 chars for richer context
- Update GPT system prompt to use all available data fields
- Add mode toggle (AI default / Algorithm fallback) in assignment preview
- Lift AI mutation to parent page for background generation persistence
- Show visual indicator on page while AI generates (spinner + progress card)
- Toast notification with "Review" action when AI completes
- Staggered reveal animation for assignment results (streaming feel)
- Fix assignment balance with dynamic penalty (25pts per existing assignment)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When closing round N, any active rounds with lower sortOrder in the
same competition are automatically closed. Each cascade closure is
recorded in DecisionAuditLog with closedBy: 'cascade' reference.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Fix assignments page header overflow on mobile (flex-wrap)
- Hide 'Back to Dashboard' button on mobile (logo tap navigates home)
- Make logo/brand text clickable to navigate to role dashboard
- Snap windowOpenAt to now when manually activating a round early
- Gate Compare Projects cards behind jury_compare_enabled setting (defaults off)
- Expose jury_compare_enabled in getFeatureFlags tRPC procedure
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Some PDFs contain \x00 null bytes in their text which PostgreSQL rejects
with "invalid byte sequence for encoding UTF8: 0x00". Sanitize extracted
text in both document-analyzer and file-content-extractor services.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
pdf-parse v2 requires DOMMatrix (browser API) which fails in Node.js.
Replaced with unpdf (serverless PDF.js build) for PDFs and mammoth for
Word .docx files. Also fixed the same broken pdf-parse usage in
file-content-extractor.ts used by AI filtering.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Introduces a document analyzer service that extracts page count (via pdf-parse),
text preview, and detected language (via franc) from uploaded files. Analysis runs
automatically on upload (configurable via SystemSettings) and can be triggered
retroactively for existing files. Results are displayed as badges in the FileViewer
and fed to AI screening for language-based filtering criteria.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Wire batchCheckRequirementsAndTransition into round activation and reopen
so pre-existing projects that already have all required docs get auto-
passed. Also adds checkDocumentCompletion endpoint for manual sweeps on
already-active rounds.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The AI assignment path was receiving project tags as flat strings, losing
the confidence scores from AI tagging. Now both the GPT path and the
fallback algorithm weight tag matches by confidence — a 0.9 tag matters
more than a 0.5 one.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add checkRequirementsAndTransition() to round-engine that checks if all
required FileRequirements for a round are satisfied by uploaded files.
When all are met and the project is PENDING/IN_PROGRESS, it auto-
transitions to PASSED. Also adds batchCheckRequirementsAndTransition()
for bulk operations.
Wired into:
- file.adminUploadForRoundRequirement (admin bulk upload)
- applicant.saveFileMetadata (applicant self-upload)
Non-fatal: failures in the check never break the upload itself.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Change AI tagging to use AI_MODELS.QUICK (gpt-4o-mini) instead of gpt-4o for
10-15x cost reduction on classification tasks
- Add openai_base_url system setting for OpenAI-compatible providers
(OpenRouter, Groq, Together AI, local models)
- Reset OpenAI client singleton when API key, base URL, or model changes
- Add base URL field to AI settings form with provider examples
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- AI Tagging: batch 10 projects per API call with 3 concurrent batches (~10x faster)
- New `tagProjectsBatch()` with `getAISuggestionsBatch()` for multi-project prompts
- Single DB query for all projects, single anonymization pass
- Compact JSON in prompts (no pretty-print) saves tokens
- AI Shortlist: run STARTUP and BUSINESS_CONCEPT categories in parallel (2x faster)
- AI Filtering: increase default parallel batches from 1 to 3 (3x faster)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Round engine: moved logAudit() calls outside $transaction blocks to prevent
FK violations from poisoning PostgreSQL transactions and rolling back status changes.
Round detail page: redesigned with Editorial Command Center aesthetic -
dark blue gradient header, colored accent stat cards, underline tab bar,
SVG readiness ring, grouped quick actions, branded progress bars and animations.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add reopenRound() to round engine (CLOSED → ACTIVE) with auto-pause of subsequent active rounds
- Add reopen endpoint to roundEngine router and UI button on round detail page
- Replace free-text MIME type input with toggle-only badge buttons in file requirements editor
- Enable refetchOnWindowFocus and shorter polling intervals for readiness checklist queries
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Per-juror cap mode (HARD/SOFT/NONE) in add-member dialog and members table
- Jury invite flow: create user + add to group + send invitation from dialog
- Per-round config: notifyOnAdvance, aiParseFiles, startupAdvanceCount, conceptAdvanceCount
- Moved notify-on-advance from competition-level to per-round setting
- AI filtering: round-tagged files with newest-first sorting, optional file content extraction
- File content extractor service (pdf-parse for PDF, utf-8 for text files)
- AI shortlist runs independently per category (STARTUP / BUSINESS_CONCEPT)
- generateAIRecommendations tRPC endpoint with per-round config integration
- AI recommendations UI: trigger button, confirmation dialog, per-category results display
- Category-aware advance dialog: select/deselect projects by category with target caps
- STAGE_ACTIVE bug fix in assignment router
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace Pipeline/Stage system with Competition/Round architecture.
New schema: Competition, Round (7 types), JuryGroup, AssignmentPolicy,
ProjectRoundState, DeliberationSession, ResultLock, SubmissionWindow.
New services: round-engine, round-assignment, deliberation, result-lock,
submission-manager, competition-context, ai-prompt-guard.
Full admin/jury/applicant/mentor UI rewrite. AI prompt hardening with
structured prompts, retry logic, and injection detection. All legacy
pipeline/stage code removed. 4 new migrations + seed aligned.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Remove RoutingRule model and routing engine (replaced by direct award assignment)
- Simplify RoutingMode enum: PARALLEL/POST_MAIN → SHARED, keep EXCLUSIVE
- Remove routing router, routing-rules-editor, and related tests
- Update pipeline, award, and notification code to remove routing references
- Seed: include all CSV entries (no filtering/dedup), AI screening handles duplicates
- Seed: fix non-breaking space (U+00A0) bug in category/issue mapping
- Stage filtering: add duplicate detection that flags projects for admin review
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Simplify pipeline list cards: whole card is clickable, remove clutter
- Add wizard edit page for existing pipelines with full state pre-population
- Extract toWizardTrackConfig to shared utility for reuse
- Rewrite predicate builder with 3 modes: Simple (sentence-style), AI (NLP), Advanced (JSON)
- Fix routing operators to match backend (eq/neq/in/contains/gt/lt)
- Rewrite routing rules editor with collapsible cards and natural language summaries
- Add parseNaturalLanguageRule AI procedure for routing rules
- Add per-category quotas to SelectionConfig and EvaluationConfig
- Add category quota UI toggles to selection and assignment sections
- Add category breakdown display to selection panel
- Add category-aware scoring to smart assignment (penalty/bonus)
- Add category-aware filtering targets with excess demotion
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- New projects (admin create, CSV import, public form) auto-assign to program's
first round (by sortOrder) when no round is specified
- Closing a FILTERING round auto-starts filtering job (configurable via
autoFilterOnClose setting, defaults to true)
- Add SUBMISSION_RECEIVED notification type for confirming submissions
- Replace separate List/Pipeline toggle with integrated pipeline view below
the sortable round list
- Add autoFilterOnClose toggle to filtering round type settings UI
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Extract observer dashboard to client component, add PDF export button
- Add PDF report generator with jsPDF for analytics reports
- Overhaul jury evaluation page with improved layout and UX
- Add new analytics endpoints for observer/admin reports
- Improve round creation/edit forms with better settings
- Fix filtering rules page, CSV export dialog, notification bell
- Update auth, prisma schema, and various type fixes
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Features implemented:
- F1: Email digest notifications with cron endpoint and per-user frequency
- F2: Jury availability windows and workload preferences in smart assignment
- F3: Round templates with save-from-round and CRUD management
- F4: Side-by-side project comparison view for jury members
- F5: Real-time voting dashboard with Server-Sent Events (SSE)
- F6: Live voting UX: QR codes, audience voting, tie-breaking, score animations
- F7: File versioning, inline preview, bulk download with presigned URLs
- F8: Mentor dashboard: milestones, private notes, activity tracking
- F9: Communication hub with broadcasts, templates, and recipient targeting
- F10: Advanced analytics: cross-round comparison, juror consistency, diversity metrics, PDF export
- F11: Applicant draft saving with magic link resume and cron cleanup
- F12: Webhook integration layer with HMAC signing, retry, and delivery logs
- F13: Peer review discussions with anonymized scores and threaded comments
- F14: Audit log enhancements: before/after diffs, session grouping, anomaly detection, retention
- F15: i18n foundation with next-intl (EN/FR), cookie-based locale, language switcher
Schema: 12 new models, field additions to User, Project, ProjectFile, LiveVotingSession, LiveVote, MentorAssignment, AuditLog, Program
New routers: roundTemplate, message, webhook (registered in _app.ts)
New services: email-digest, webhook-dispatcher
New cron endpoints: /api/cron/digest, /api/cron/draft-cleanup, /api/cron/audit-cleanup
New API routes: /api/live-voting/stream (SSE), /api/files/bulk-download
All features are admin-configurable via SystemSettings or per-model settingsJson fields.
Docker build verified successfully.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Security (Critical/High):
- Fix path traversal bypass in local storage provider (path.resolve + prefix check)
- Fix timing-unsafe HMAC comparison (crypto.timingSafeEqual)
- Add auth + ownership checks to email API routes (verify-credentials, change-password)
- Remove hardcoded secret key fallback in local storage provider
- Add production credential check for MinIO (fail loudly if not set)
- Remove DB error details from health check response
- Add stricter rate limiting on application submissions (5/hour)
- Add rate limiting on email availability check (anti-enumeration)
- Change getAIAssignmentJobStatus to adminProcedure
- Block dangerous file extensions on upload
- Reduce project list max perPage from 5000 to 200
Query Optimization:
- Optimize analytics getProjectRankings with select instead of full includes
- Fix N+1 in mentor.getSuggestions (batch findMany instead of loop)
- Use _count for files instead of fetching full file records in project list
- Switch to bulk notifications in assignment and user bulk operations
- Batch filtering upserts (25 per transaction instead of all at once)
UI/UX:
- Replace Inter font with Montserrat in public layout (brand consistency)
- Use Logo component in public layout instead of placeholder
- Create branded 404 and error pages
- Make admin rounds table responsive with mobile card layout
- Fix notification bell paths to be role-aware
- Replace hardcoded slate colors with semantic tokens in admin sidebar
- Force light mode (dark mode untested)
- Adjust CardTitle default size
- Improve muted-foreground contrast for accessibility (A11Y)
- Move profile form state initialization to useEffect
Code Quality:
- Extract shared toProjectWithRelations to anonymization.ts (removed 3 duplicates)
- Remove dead code: getObjectInfo, isValidImageSize, unused batch tag functions, debug logs
- Remove unused twilio dependency
- Remove redundant email index from schema
- Add actual storage object deletion when file records are deleted
- Wrap evaluation submit + assignment update in
- Add comprehensive platform review document
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add try-catch around OpenAI API call to catch model errors
- Provide clearer error messages when model name is invalid
- Handle empty response cases with helpful feedback
- Suggest checking AI Configuration settings when model fails
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Log each project being processed with timing
- Log progress every 10 projects with time estimates
- Log final completion stats
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Redesign AI Tagging dialog with scope selection (Round vs Edition)
- Add visual progress indicator during AI processing
- Display result stats (tagged/skipped/failed) after completion
- Add batchTagProgramProjects endpoint for edition-wide tagging
- Fix getFilterOptions to include program.id for filtering
- Improve error handling with toast notifications
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Fall back to ai_enabled setting if ai_tagging_enabled not set
- Check both tags array and projectTags relationship for untagged projects
- Add detailed logging for debugging
- Show errors in toast instead of just counts
- Handle "no projects to tag" case with appropriate message
- Add early checks for OpenAI config and available tags
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add AssignmentJob model to track AI assignment progress
- Create startAIAssignmentJob mutation for background processing
- Add getAIAssignmentJobStatus query for polling progress
- Update AI assignment service with progress callback support
- Add progress bar UI showing batch/project processing status
- Add toast notifications for job completion/failure
- Add AI_SUGGESTIONS_READY notification type
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add minAssignmentsPerJuror and maxAssignmentsPerJuror fields to Round model
- Update assignment router:
- Calculate effective max from user override or round default
- Add forceOverride parameter for manual assignment beyond limits
- Update getSuggestions to use round constraints with min target bonus
- Update getAISuggestions to pass constraints to AI service
- Update AI assignment service:
- Add minAssignmentsPerJuror to constraints interface
- Update fallback algorithm with under-min bonus scoring
- New score weights: 50% expertise, 30% load, 20% under-min bonus
- Update round router:
- Add new constraint fields to create/update schemas
- Add validation for min <= max constraint
- Update admin UI:
- Add min/max constraint fields to round edit page
- Remove hardcoded maxPerJuror from assignments page
- Add migration files for production deployment:
- User.bio field for judge/mentor profiles
- Round assignment constraint fields
Constraint hierarchy:
1. User.maxAssignments (if set) overrides round default
2. Round.maxAssignmentsPerJuror is the default cap
3. Admin can force-override any limit with confirmation
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add bio field to User model for judge/mentor profile descriptions
- Add bio step to onboarding wizard (optional step with 500 char limit)
- Enhance smart assignment to match judge bio against project description
- Uses keyword extraction and Jaccard-like similarity scoring
- Only applies if judge has a bio (no penalty for empty bio)
- Max 15 points for bio match on top of existing scoring
- Fix geographic distribution query to use round relation for programId
- Update score breakdown: tags (40), bio (15), workload (25), country (15)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
