Team Contribution Report — Labs 1–9 (Through Phase 3)

Project: LexGuard — Neuro-Symbolic Compliance Auditor for Contract Risk Analysis Course: CS 5542 — Big Data Analytics & Applications, UMKC, Spring 2026 GitHub: https://github.com/JoeDoan/Lab9_BigData

---

Phase 3 — Team Contribution Table

Team Member	Role	Phase 3 Contributions	%
Joe Doan	Data Pipeline & Adaptation Lead	BERT vs LLM evaluation (evaluate_e2e.py), full-doc LLM extraction pipeline (extract_risk_clauses_llm, extract_contract_brief), Snowflake chat persistence (chat_history.py), dark/light theme toggle, chat history UI with delete & LLM-generated titles, Phase 3 report	30%
Manan Koradiya	Agent Architect & Integrator	app.py UI redesign (glassmorphism CSS, chat interface), RAG fallback enhancement (tools.py), end-to-end system integration, reasoning panels and query history sidebar	25%
Aditya Naredla	Storage & Evaluation Engineer	PEFT training notebook (LexGuard_PEFT_Training.ipynb), monitor.py module, live analytics dashboard, HuggingFace Hub adapter upload	25%
Ruixuan Hou	Reproducibility Lead	requirements.txt, Dockerfile, .streamlit/config.toml, reproduce.sh, REPRO_AUDIT.md, RUN.md setup instructions, system status panel	20%
Total			100%

---

Phase 3 — Key Technical Decisions

1. BERT → Full-Document LLM Extraction

• Fine-tuned BERT QA model (doandune/LexGuard-CUAD-BERT) achieved only 53.8% accuracy with ~0% recall on 12 risk clause types.
• Root cause: BERT's 512-token window misses clauses spanning multiple paragraphs.
• Decision: Replaced with Gemini 2.5 Flash full-document extraction (86.3% accuracy), passing up to 200K characters directly to the LLM.

2. Chunking + RAG → Direct Full-Document Input

• Evaluated hybrid retrieval (FAISS + BM25 + cross-encoder reranking) with document chunking.
• Chunking fragmented important clause context, lowering extraction accuracy.
• Decision: Production pipeline now feeds the entire document directly to Gemini, leveraging its 1M-token context window.

3. Snowflake Chat Persistence

• Added CHAT_SESSIONS and CHAT_MESSAGES tables with annotation metadata serialization (JSON).
• LLM-generated session titles, delete functionality, and full session restore including expandable source annotations.

---

System Architecture (Phase 3 Production)

User (Streamlit UI — Dark/Light Theme)
        ↓
  [File Upload: PDF/TXT]
        ↓
  PyMuPDF Text Extraction (Full Document)
        ↓
  ┌─────────────────────────────────────┐
  │ PRIMARY PATH: Full-Doc LLM         │
  │   • Risk Audit (200K chars → Gemini)│
  │   • Metadata Brief (8 entities)     │
  │   • General Q&A (50K chars)         │
  └─────────────────────────────────────┘
        ↓
  Gemini 2.5 Flash Response
  + Expandable Source Annotations
        ↓
  Snowflake Persistence
  (CHAT_SESSIONS + CHAT_MESSAGES + METADATA)

---

Lab 9 — Team Contribution Table

Team Member	Role	Lab 9 Contributions	%
Joe Doan	Data Pipeline & Adaptation Lead	Structured execution traces in agent.py and adapted_agent.py, timed tool calls, trace-based debug logging, LAB9_REPORT.md	30%
Manan Koradiya	Agent Architect & Integrator	Complete app.py UI redesign with premium dark theme, glassmorphism CSS, chat interface, reasoning panels, query history sidebar, error handling	25%
Aditya Naredla	Storage & Evaluation Engineer	monitor.py module (QueryMetrics + MetricsCollector), live analytics dashboard in sidebar, per-pipeline latency comparison	25%
Ruixuan Hou	Reproducibility Lead	requirements.txt, .streamlit/config.toml, Dockerfile, deployment configuration, system status panel	20%
Total			100%

---

Lab 8 — Team Contribution Table

Team Member	Role	Lab 8 Contributions	%
Joe Doan	Data Pipeline & Adaptation Lead	Instruction dataset generation (generate_dataset.py), adapted_agent.py full pipeline, Colab FastAPI server debugging, prompt format fix, response parsing, EVALUATION.md	30%
Manan Koradiya	Agent Architect & Integrator	Streamlit baseline vs. adapted toggle (app.py), RAG fallback enhancement (tools.py), end-to-end system integration	25%
Aditya Naredla	Storage & Evaluation Engineer	Domain task definition, model selection (Llama-3), PEFT training notebook (LexGuard_PEFT_Training.ipynb), HuggingFace Hub adapter upload, evaluation design	25%
Ruixuan Hou	Reproducibility Lead	reproduce.sh Lab 8 updates, new smoke tests for adapted pipeline, REPRO_AUDIT.md non-determinism documentation, RUN.md setup instructions	20%
Total			100%

---

Deliverables Summary

Deliverable	File	Status
Phase 3 Report	Phase_3_Report_LexGuard.docx	✅ Complete
Full-Doc LLM Extraction	tools.py (extract_risk_clauses_llm, extract_contract_brief)	✅ Production
Chat Persistence	chat_history.py	✅ Snowflake-backed
Dark/Light Theme	app.py (CSS variables + toggle)	✅ Deployed
BERT Evaluation	evaluate_e2e.py	✅ 53.8% → deprecated
Premium Streamlit UI	app.py	✅ Dark theme + glassmorphism
Monitoring Module	monitor.py	✅ QueryMetrics + Analytics
Structured Traces	agent.py, adapted_agent.py	✅ Timed tool calls
Deployment Config	Dockerfile, .streamlit/config.toml	✅ Docker + Theme
Dependencies	requirements.txt	✅ Pinned versions
Development Report	LAB9_REPORT.md	✅ Complete
Individual Reports	CONTRIBUTION_*.md	✅ All 4 members