V4.0.0 - Final Release

Final Release Notes

Date: 2026-05-24

Release Theme

The final release turns the Group104 TA Recruitment System from a minimal
Swing/CSV workflow into a more complete local recruitment prototype with role
separation, profile-aware review, safer Admin oversight, and optional local AI
matching.

The main release message is:

• stable Java desktop app
• transparent CSV persistence
• complete TA/MO/Admin demo workflow
• local semantic AI when available
• deterministic Java fallback when the AI server is unavailable

Headline Features

Local semantic AI matching

• Uses sentence-transformers/all-MiniLM-L6-v2.
• Encodes TA profile text and job post text into 384-dimensional vectors.
• Uses cosine similarity to produce an AI match score.
• Powers TA job ranking in Discover Jobs.
• Powers MO applicant ranking in applicant review.
• Runs locally without cloud APIs or paid keys.

Reliable fallback path

• build_and_run_mac.sh tries to start the semantic server.
• If setup fails, if the model is unavailable, or if the server does not become healthy, Java still starts.
• Java then uses AiMatchService, the built-in keyword matcher.
• This keeps the final demo safe on machines without Python model dependencies.

Mac setup cache reuse

• setup_ai_mac.sh reuses ai_venv/ when present.
• It reuses installed sentence-transformers when present.
• It reuses .ai_models/ and the model-ready marker when present.
• Repeated runs should not redownload weights after the cache is ready.

Role workflow completion

• TA: profile, job discovery, search/filter, AI ranking, application, status tracking, withdraw.
• MO: profile, job posting, job editing, applicant ranking, review context, decisions, messaging.
• ADMIN: dashboard, school/college management, safe correction for orphan application rows.

Final Release Scope

Implemented:

• password-based login and registration for TA, MO, and ADMIN
• shared profile editing
• Admin-managed school/college options
• richer job posts with description, required skills, and tags
• TA job search/filter and AI-based default sorting
• MO posted-job management and editing
• MO applicant review with application note, profile context, job context, messaging, and AI score
• Admin dashboard and safe correction
• macOS one-command build with optional semantic AI startup
• local model-cache strategy and fallback documentation
• README research-style generated figures

Out of scope:

• CV upload or CV parsing
• cloud LLM integration
• password reset/recovery
• formal deadlines, pay, workload-hour, or timetable scheduling
• arbitrary Admin CSV editing
• committed model weights or Python virtual environments

Validation Summary

Final documentation and launch-path checks:

bash -n build_and_run_mac.sh setup_ai_mac.sh
python3 -m py_compile tools/ai_semantic_match.py tools/ai_semantic_server.py
find codes/src -name '*.java' -print0 | xargs -0 javac -d /tmp/group104-build-check
git diff --check

Previously executed final-sprint checks also covered:

• generated large TA/MO test data
• 50-job Jobs Panel stress data
• AI job sort behaviour for TAs
• AI applicant sort behaviour for MOs
• semantic client response parsing
• fallback behaviour when semantic server is unavailable
• macOS script startup with copied local model cache

Demo Talking Points

1. The project remains simple to run: Java 21 plus one launcher script.
2. The app does not require a database server; all runtime data stays in CSV.
3. The AI layer is real local embedding similarity, not a fake hard-coded score.
4. The AI layer is safe for demos because Java fallback keeps the app usable.
5. The role workflows are explainable: TAs discover and apply, MOs review and decide, Admin monitors and corrects narrow data issues.
6. The final release is aligned with the project process: small increments, manual testing, issue-driven fixes, and stable main.