# Getting Started: HVAC Blueprint Analyzer Research Plan

**Date:** June 2026  
**Status:** Ready to begin Phase 3 (Experimental Evaluation)

---

## What Just Happened

You now have a **complete research-to-publication plan** for converting your HVAC Blueprint Analyzer CLI tool into a publishable paper. Here's what was created:

### 📚 Six Planning Documents (Total: ~70 KB)

| Document | Purpose | Key Output |
|----------|---------|-----------|
| **RESEARCH_FOUNDATION.md** | Literature review & positioning | 100+ citations, novelty statement |
| **RESEARCH_UPDATE.md** | Newly discovered papers | 7 papers grounding your methodology |
| **CODE_IMPROVEMENTS.md** | Research-backed enhancements | 9 improvements, code examples, effort estimates |
| **INTEGRATED_PLAN.md** | Master plan | 18 tasks, 3.5-month timeline, dependencies |
| **PAPER_OUTLINE.md** | Paper template | Section-by-section structure, placeholders |
| **PUBLICATION_ROADMAP.md** | Execution phases | 7 phases from research to submission |

### ✅ 18 Tasks Created (With Dependencies)

All tasks are in the Cowork task system with clear dependencies:

```
Phase 3: Baseline Metrics (#1)
    ↓
Code Improvements (#2-9)
    ↓
Testing & Validation (#10)
    ↓
Paper Writing (#11-16)
    ↓
Open Source + Submission (#17-18)
```

### 🔗 Updated Main README

The project's `README.md` now links to all research documents and explains the publication roadmap upfront.

---

## Your Next Steps (This Week)

### Priority 1: Understand the Plan

**Read these in order (30 min total):**

1. ⏳ **[INTEGRATED_PLAN.md](INTEGRATED_PLAN.md)** — 10 min read
   - Understand the 18-task workflow
   - See timelines and milestones
   - Identify any blockers upfront

2. ⏳ **[CODE_IMPROVEMENTS.md](CODE_IMPROVEMENTS.md)** — Priority 1 section only (10 min)
   - Tasks #2-4 (Rotation preprocessing, Confidence tracking, Equipment lookup)
   - Understand effort estimates and code examples

3. ⏳ **[RESEARCH_FOUNDATION.md](RESEARCH_FOUNDATION.md)** — Section 9 (10 min)
   - Your paper's positioning vs. AHMsys, Bergström, Zhang, etc.
   - Why your approach is novel

### Priority 2: Assess Your Current Data

**Before starting Phase 3, answer these questions:**

1. **Do you have 10-30 diverse blueprints?** (Different architects, eras, quality levels)
   - If no: Source them this week
   - If yes: Select representative 10-20 for baseline evaluation

2. **Do you have manual annotations for ground truth?** (Correct schedules + tag locations)
   - If no: Plan to manually annotate 30 min per blueprint
   - If yes: Format as JSON for evaluation script

3. **Is your equipment reference sheet complete?** (Mitsubishi_Daikin_Indoor_HVAC_Units_v1.xlsx)
   - Check: What % of model numbers in your data match existing prefixes?
   - Gap = opportunity for Task #9 (model validation improvement)

4. **Gemini API cost:** How much budget do you have?
   - Tiered approach targets 82% cost reduction vs. Gemini-only
   - Verify your .env has GEMINI_API_KEY set

---

## Understanding the Research Grounding

Your code improvements are **not arbitrary**. Each is grounded in published 2024-2025 research:

### Priority 1 Tasks

**Task #2 (Rotation Preprocessing):**
- Research: Bergström et al. (2025) — "Optimizing Text Recognition in Mechanical Drawings"
- Claim: 15-20% improvement in OCR success on rotated text
- Implementation: Use Tesseract OSD (orientation detection) + CLAHE enhancement

**Task #3 (Confidence Tracking):**
- Research: Chen et al. (2024) + Devulapalli & Kumar (2024)
- Claim: Confidence-based routing beats single-stage methods
- Implementation: ExtractionResult class with escalation thresholds

**Task #4 (Equipment Reference Lookup):**
- Research: AHMsys (2024) + Zou et al. (2023)
- Claim: 10-15% recovery of missing unit types via prefix matching
- Implementation: Longest-prefix matching against Excel sheet

### Why This Matters for Your Paper

When you write Section 5 (Evaluation) of the paper, you'll be able to say:

> "Prior work shows that rotation-aware preprocessing improves OCR on mechanical drawings by 15-20% (Bergström 2025). Our implementation achieves [X]% improvement on the same baseline dataset, validating the approach."

This is **publishable research**, not just code improvements.

---

## Timeline & Effort Summary

| Phase | Timeline | Effort | Key Deliverable |
|-------|----------|--------|-----------------|
| **Phase 3** | Weeks 1-2 | 10-15 hrs | baseline_metrics.json |
| **Code Improvements** | Weeks 2-4 | ~21 hrs | 9 new/modified modules |
| **Testing & Validation** | Weeks 4-5 | 5-10 hrs | improvement_metrics.json |
| **Paper Writing** | Weeks 5-8 | 30-40 hrs | Full 8-page draft |
| **Review & Polish** | Weeks 8-9 | 10-15 hrs | Polished draft |
| **Open Source** | Week 9 | 4-6 hrs | GitHub repo |
| **Submission** | Week 10 | 2-4 hrs | Paper submitted |

**Total: ~85-105 hours over 10 weeks** (roughly 2.5 hours/day if working full-time on this)

---

## Recommended Starting Point

### This Week: Start Task #1 (Phase 3 Baseline Metrics)

**What to do:**

1. **Collect 10-20 representative blueprints**
   - Different architects/regions if possible
   - Mix of quality levels (clear scans + degraded scans)
   - Save to: `raw_pdfs/` folder

2. **Create ground truth annotations** (JSON format)
   ```json
   {
     "file": "building_name_mechanical_100%.pdf",
     "schedule_page": 29,
     "units": [
       {"tag": "AC-A", "qty": 2, "type": "Wall Mounted", "model": "FXAQ09PVJU", "brand": "Mitsubishi"},
       {"tag": "AC-B", "qty": 1, "type": "Ceiling Cassette", "model": "PLFY12BAC", "brand": "Daikin"}
     ],
     "floor_plans": [
       {"page": 3, "floor": "01", "detected_units": [{"tag": "AC-A", "coords": [100, 200]}, ...]}
     ]
   }
   ```

3. **Run current pipeline on these blueprints**
   ```powershell
   python analyze_blueprint.py "raw_pdfs/building.pdf" --yes
   ```

4. **Collect baseline metrics** (save to `baseline_metrics.json`)
   - How many tags extracted at Tier 1, 2, 3, 4?
   - Average confidence per tier
   - Total Gemini API calls
   - Latency per tier

5. **Compare vs. ground truth**
   - Which extractions are wrong?
   - Where did we escalate to Tier 4 unnecessarily?

---

## Key Files You'll Be Working With

### Code Files to Modify/Create:

```
hvac_pipeline/
├── analyze_blueprint.py              # Main entry (modify for metrics collection)
├── extract_schedule_robust.py         # Schedule extraction (add confidence tracking)
├── reconcile_schedule_vs_plans.py     # Reconciliation (add cross-page boost)
├── preprocess_for_ocr.py             # NEW: Rotation preprocessing
├── equipment_reference.py             # NEW: Reference sheet lookup
├── page_classifier.py                 # NEW: Page classification
└── metrics.py                         # NEW: Metrics collector
```

### Documentation Files (For Reference):

```
├── RESEARCH_FOUNDATION.md             # 100+ citations, positioning
├── CODE_IMPROVEMENTS.md               # 9 enhancements with code examples
├── INTEGRATED_PLAN.md                 # 18-task master plan
├── PAPER_OUTLINE.md                   # Paper template with sections
└── PUBLICATION_ROADMAP.md             # 7-phase timeline
```

---

## Questions? Check These First

**"How do I understand what my novelty is?"**
→ Read: RESEARCH_FOUNDATION.md, Section 9 (Novelty & Positioning)

**"What code should I write first?"**
→ Read: CODE_IMPROVEMENTS.md, Priority 1.2 (Confidence tracking unblocks metrics)

**"What goes in the paper?"**
→ Read: PAPER_OUTLINE.md (section-by-section template with placeholders)

**"What's the complete timeline?"**
→ Read: INTEGRATED_PLAN.md (18 tasks with dependencies, 10-week schedule)

**"What research papers should I cite?"**
→ Read: RESEARCH_FOUNDATION.md, Section 11 (organized bibliography)

---

## Success Metrics

### By End of Week 2:
- ✅ Task #1 complete: baseline_metrics.json showing Tier 1-4 performance
- ✅ Ground truth annotations ready for 10-20 blueprints
- ✅ You understand your novelty vs. 7 key papers

### By End of Week 4:
- ✅ Tasks #2-9 implemented (all code improvements)
- ✅ Code tested on same dataset as Task #1

### By End of Week 5:
- ✅ improvement_metrics.json showing before/after comparison
- ✅ Data ready to populate paper's Evaluation & Results sections

### By End of Week 8:
- ✅ Full 8-page draft (all sections written)
- ✅ Ready for internal review

### By End of Week 10:
- ✅ Paper submitted to target venue (ICDAR or alternative)
- ✅ Code released on GitHub

---

## Final Checklist Before Starting

- [ ] Read INTEGRATED_PLAN.md (understand the 18 tasks)
- [ ] Read CODE_IMPROVEMENTS.md Priority 1 section (understand effort)
- [ ] Read RESEARCH_FOUNDATION.md Section 9 (understand your novelty)
- [ ] Verify GEMINI_API_KEY is set in .env
- [ ] Have 10-20 representative blueprints ready
- [ ] Know how you'll create ground truth annotations
- [ ] Review PAPER_OUTLINE.md to see what results you'll need

---

## You're Ready!

You have:
- ✅ Complete research foundation (100+ citations)
- ✅ Clear code improvements (9 tasks, 21.5 hours)
- ✅ Paper template (section-by-section outline)
- ✅ Master plan (18 interdependent tasks)
- ✅ Timeline (3.5 months to submission)

**Next action: Start Phase 3 (Task #1).**

Gather baseline metrics on your representative blueprint dataset. Everything else depends on this.

Good luck! 🚀

---

*For questions or clarifications, refer to the appropriate planning document listed above. All documents are in the hvac_project/ root folder.*