diff --git a/work/document_AI_analysis.md b/work/document_AI_analysis.md
new file mode 100644
index 0000000..6213333
--- /dev/null
+++ b/work/document_AI_analysis.md
@@ -0,0 +1,222 @@
+Here’s a **no-nonsense one-pager** you can use to pitch law firms, focusing on **immediate time savings** and **risk reduction** (no tech jargon):
+
+---
+
+# **🚀 Auto-Redline: Cut Contract Review Time by 80%**  
+### **AI-Powered Redlining for Mid-Sized Law Firms**  
+
+**What It Does**:  
+- **Instantly redlines** opponent drafts (Word/PDF) against your firm’s playbook.  
+- **Flags hidden risks** (e.g., auto-renewals, unusual liability caps).  
+- **Suggests pre-approved clauses** with one-click replacement.  
+
+**How It Works**:  
+1. **Upload** a contract (drag & drop into Word/Outlook).  
+2. **AI highlights** problematic terms + suggests fixes.  
+3. **You review** and click "Accept" or "Revise."  
+
+---
+
+### **Why Firms Love It**  
+✅ **Saves 15+ hours/week** per lawyer (no more manual redlining).  
+✅ **Catches sneaky terms** even partners miss (e.g., "This indemnification clause is 3x broader than market").  
+✅ **Works offline**—no data leaves your servers.  
+
+---
+
+### **Use Cases**  
+- **NDAs**: Review in **5 minutes** vs. 2 hours.  
+- **Leases**: Auto-flag **unusual covenants**.  
+- **Employment Agreements**: Enforce **firm-approved templates**.  
+
+---
+
+### **Competitive Edge**  
+| Feature | Legacy Tools | Auto-Redline |  
+|---------|-------------|--------------|  
+| **Handles scanned PDFs/tables** | ❌ | ✅ |  
+| **Learns your firm’s preferences** | ❌ | ✅ |  
+| **One-click clause replacement** | ❌ | ✅ |  
+
+---
+
+### **Pricing**  
+- **$499/month** (unlimited contracts, per user).  
+- **7-day free trial** (no credit card needed).  
+
+---
+
+### **Get Started**  
+📞 **Call**: [Your Number]  
+🌐 **Demo**: [Your Website]  
+
+> *"We cut NDA review time from 2 hours to 10 minutes. Game-changer!"*  
+> — [Law Firm Name], [Title]  
+
+---
+
+### **FAQ**  
+**Q: Is this secure?**  
+A: Yes—runs 100% on your computers. No cloud required.  
+
+**Q: How long to set up?**  
+A: 5 minutes. We pre-load your clause library.  
+
+**Q: What if we hate it?**  
+A: Cancel anytime. No contracts.  
+
+---
+
+**Design Notes**:  
+- Use **bold, clean headings** (lawyers skim).  
+- Include a **testimonial** (social proof).  
+- **Avoid tech terms** ("AI" → "auto-redline").  
+
+Want this as a **PDF template**? Happy to customize it further! This pitch works because it **solves a daily pain point**—not because the tech is "cool."
+
+Here's a **crystal-clear tech stack breakdown** with deployment milestones, designed for **scalability** and **mid-firm adoption**. We'll focus on **minimum viable components** that deliver maximum value fast:
+
+---
+
+### **⚙️ Core Tech Stack**
+#### **1. Document Ingestion & Parsing Layer**
+| Component | Purpose | Key Features | Alternatives |
+|-----------|---------|--------------|--------------|
+| **Docling** | Parse complex contracts (PDFs, scans, tables) | - Layout-aware extraction<br>- OCR for scans<br>- Table/formula detection | Adobe PDF Extract API (costly) |
+| **Apache Tika** (fallback) | Extract text from uncommon formats | - Supports 1,000+ file types | - |
+
+**Deployment Goal**:  
+- Ingest 95% of contract types (PDF, DOCX, scans) with **>90% accuracy**.
+
+---
+
+#### **2. Data Storage & Query Layer**
+| Component | Purpose | Key Features | Alternatives |
+|-----------|---------|--------------|--------------|
+| **DuckDB** | OLAP analytics on contracts | - SQL queries on JSON/Parquet<br>- Client-side processing | Snowflake (overkill) |
+| **Parquet Files** | Store processed contracts | - Columnar efficiency<br>- Versioning via Delta Lake | MongoDB (less performant) |
+
+**Deployment Goal**:  
+- Execute **10,000+ clause searches/sec** on a laptop.
+
+---
+
+#### **3. AI/ML Layer**
+| Component | Purpose | Key Features | Alternatives |
+|-----------|---------|--------------|--------------|
+| **Haystack** | Redlining & Q&A | - Pre-built RAG pipelines<br>- Local LLM support | LangChain (more dev-heavy) |
+| **all-MiniLM-L6-v2** | Embeddings | - Lightweight<br>- 384-dim vectors | OpenAI embeddings (cloud) |
+| **Phi-3-small** (optional) | Local LLM | - 4-bit quantized<br>- Runs on CPU | LLaMA-3 (larger) |
+
+**Deployment Goal**:  
+- **Redline a 50-page contract in <30 sec** on a MacBook Pro.
+
+---
+
+#### **4. Integration Layer**
+| Component | Purpose | Key Features | Alternatives |
+|-----------|---------|--------------|--------------|
+| **FastAPI** | Backend API | - Python-native<br>- Swagger docs | Flask (less async) |
+| **Microsoft Word Add-in** | User interface | - Office JS API<br>- Track changes integration | None (critical for adoption) |
+
+**Deployment Goal**:  
+- **One-click redline** from Word’s ribbon toolbar.
+
+---
+
+### **🚀 Deployment Phases**
+#### **Phase 1: Local MVP (4 Weeks)**
+- **Target**: Single-lawyer usability  
+- **Deliverables**:  
+  1. Docling → DuckDB pipeline ingesting **PDFs + DOCX**.  
+  2. Haystack RAG answering **"Show indemnification clauses"**.  
+  3. **Word Add-in MVP** (highlight clauses only).  
+
+#### **Phase 2: Firm-Wide (8 Weeks)**
+- **Target**: 5-user pilot  
+- **Deliverables**:  
+  1. **Playbook integration** (pre-approved clauses).  
+  2. **Batch processing** (upload 100+ contracts).  
+  3. **Basic analytics dashboard** (DuckDB + Plotly).  
+
+#### **Phase 3: Enterprise (12+ Weeks)**
+- **Target**: 50+ users  
+- **Deliverables**:  
+  1. **Self-learning** (auto-updates playbooks).  
+  2. **Opponent profiling** ("Firm X always hides arbitration").  
+  3. **SOC-2 compliance**.  
+
+---
+
+### **🔧 Developer Cheatsheet**
+#### **1. Docling → DuckDB Flow**
+```python
+# Parse contract
+from docling import DoclingDocument
+doc = DoclingDocument("contract.pdf")
+
+# Convert to DuckDB-ready JSON
+import json
+with open("contract.json", "w") as f:
+    json.dump(doc.to_dict(), f)
+
+# Query in DuckDB
+import duckdb
+duckdb.sql("""
+  SELECT 
+    text, 
+    meta->>'parties' AS parties
+  FROM read_json('contract.json')
+  WHERE text LIKE '%indemnification%'
+""")
+```
+
+#### **2. Haystack Redlining**
+```python
+from haystack import Pipeline
+from haystack.nodes import EmbeddingRetriever, FARMReader
+
+# Load pre-built index
+retriever = EmbeddingRetriever(embedding_model="all-MiniLM-L6-v2")
+reader = FARMReader(model_name_or_path="deepset/roberta-base-squad2")
+
+# Build pipeline
+pipe = Pipeline()
+pipe.add_node(component=retriever, name="Retriever", inputs=["Query"])
+pipe.add_node(component=reader, name="Reader", inputs=["Retriever"])
+
+# Run
+results = pipe.run(query="Find all termination clauses")
+```
+
+#### **3. Word Add-in (Simplified)**
+```javascript
+// Office JS API (Word Add-in)
+function highlightClauses() {
+  Word.run(async (context) => {
+    const clauses = await getAIResults(); // Call FastAPI
+    clauses.forEach(clause => {
+      context.document.getSearch(clause.text).highlight("yellow");
+    });
+    await context.sync();
+  });
+}
+```
+
+---
+
+### **📊 Key Metrics to Track**
+| Metric | Tool | Target |
+|--------|------|--------|
+| **Parse Accuracy** | Docling logs | >90% clauses correct |
+| **Redline Speed** | Haystack timers | <30 sec/contract |
+| **User Adoption** | Word Add-in telemetry | 80% weekly active users |
+
+---
+
+### **💡 Pro Tips**
+1. **Start with NDAs** (simple, high-volume).  
+2. **Use DuckDB’s JSON streaming** for >1GB contracts.  
+3. **Cache embeddings** to speed up repeat queries.  
+
+Need **detailed deployment checklists** or **sample contracts** to test with? Happy to share! This stack is battle-ready for mid-sized firms.
\ No newline at end of file