# HVAC Blueprint Analyzer: Publication Roadmap

## Status

**Project:** Converting HVAC Blueprint Analyzer from CLI tool → research paper  
**Created:** June 2026  
**Next Review:** [You decide]

---

## Phase 1: Research Foundation ✅ COMPLETE

**Deliverables:**
- ✅ `RESEARCH_FOUNDATION.md` — Comprehensive literature review organized by topic
  - Vision-language models & document understanding
  - OCR & document analysis
  - Table & structured data extraction
  - Floor plan & architectural analysis
  - HVAC-specific work
  - Your positioning vs. prior work
  - Suggested paper structure & keywords
  - Full bibliography (100+ citations)

**Next Step:** Review the foundation; ensure all citations still feel relevant as you write.

---

## Phase 2: Paper Outline & Structure (READY FOR YOUR INPUT)

**Deliverables:**
- ✅ `PAPER_OUTLINE.md` — Detailed section-by-section outline
  - Abstract template
  - Introduction frame (industry problem, why existing solutions fail, your approach)
  - Related work structure (pull from RESEARCH_FOUNDATION.md)
  - Problem formulation & data model
  - Methodology (full system architecture, each tier explained)
  - Evaluation plan (dataset, metrics, baselines, ablation)
  - Results sections (placeholders for your numbers)
  - Discussion, limitations, future work
  - Appendices

**What You Need To Do (Phase 2):**
1. **Fill in the [Subsection content to write] placeholders** with actual prose
2. **Gather your experimental results:**
   - Performance metrics (Tier 1-4 effectiveness, end-to-end accuracy)
   - Dataset size and characteristics
   - Baseline comparisons (Tesseract-only vs. Gemini-only vs. your tiered approach)
   - Cost analysis (API calls per PDF, latency breakdown)
3. **Create figures/tables:**
   - System architecture diagram
   - Tier utilization bar chart (% of blueprints resolved at each tier)
   - Performance comparison table
   - Cost-accuracy tradeoff analysis
4. **Assemble supplementary materials:**
   - Sample extracted data (JSON)
   - Ground truth annotations
   - Failure case analysis (3-5 challenging blueprints)
   - Sample floor plan annotations with YOLO labels

---

## Phase 3: Experimental Evaluation (CRITICAL)

**What You Need to Measure:**

### 3.1 Dataset Assembly
- [ ] Gather representative sample of blueprints (10-30 PDFs minimum; 50+ if possible)
- [ ] Document diversity: architect, era, quality level, HVAC brand
- [ ] Manually annotate ground truth for each blueprint:
  - Extract correct schedule by hand
  - Annotate floor plan images with unit tag bounding boxes
  - Record any ambiguities or errors in source blueprint

### 3.2 Quantitative Evaluation
- [ ] Run each tier independently on test set; measure performance
  - Tier 1 (pdfplumber): # tags extracted, confidence
  - Tier 2 (spatial text): # tags extracted, accuracy
  - Tier 3 (Tesseract): # tags extracted, confidence, OCR errors
  - Tier 4 (Gemini): # tags extracted, accuracy, cost
- [ ] Measure end-to-end accuracy (all fields correct: tag, qty, type, model, brand)
- [ ] Measure per-field accuracy (separate metrics for each column)
- [ ] Floor plan detection (YOLO): precision, recall, mAP
- [ ] Confidence calibration: do predicted scores match actual accuracy?

### 3.3 Comparison to Baselines
- [ ] Tesseract-only baseline: run Tier 3 alone, report performance
- [ ] Gemini-only baseline: skip local tiers, go straight to Tier 4
- [ ] Manual extraction baseline: time + cost (if resources permit)
- [ ] Statistical significance tests (confidence intervals, p-values)

### 3.4 Ablation Study
- [ ] Performance if you remove Tier 1 → what's the impact?
- [ ] Performance if you remove Tier 2 → what's the impact?
- [ ] Performance if you remove Tier 3 → what's the impact?
- [ ] Sensitivity to confidence thresholds (vary tier fallback triggers)
- [ ] Sensitivity to Tesseract PSM modes (PSM 4 vs. PSM 6)

### 3.5 Cost Analysis
- [ ] Count actual Gemini API calls on test set
- [ ] Measure latency per tier
- [ ] Estimate cost per blueprint at different accuracy targets
- [ ] Compare cost of tiered approach vs. Gemini-only

### 3.6 Error Analysis
- [ ] Categorize failure modes:
  - OCR errors (unrecognizable characters)
  - Parsing errors (regex failed to extract valid tag)
  - Semantic errors (wrong unit type assigned)
  - Hallucinations (Gemini invented non-existent units)
  - Incomplete extractions (missing some units)
- [ ] Create case studies for 3-5 most difficult blueprints
- [ ] Identify systematic patterns (e.g., specific fonts, layouts that fail)

---

## Phase 4: Paper Writing (DRAFT → REVISION)

**Timeline Estimate:** 4-8 weeks of focused writing

### 4.1 Writing Strategy

1. **Write sections in order of data dependency:**
   - Methodology (no data needed; foundational)
   - Evaluation (describe what you measured; cite experimental setup)
   - Results (present data; create tables/figures)
   - Discussion (interpret results; relate to prior work)
   - Related Work (synthesize RESEARCH_FOUNDATION.md citations)
   - Introduction (orient reader to problem; promise contributions)
   - Abstract (write last, after results finalized)

2. **Suggested page budget (IEEE 8-page format):**
   - Abstract: 0.25 pages
   - Introduction: 1.5 pages
   - Related Work: 2 pages
   - Methodology: 2 pages
   - Evaluation: 0.75 pages
   - Results: 1 page
   - Discussion & Future Work: 0.5 pages

3. **Tools for writing:**
   - Overleaf (free IEEE template available)
   - MS Word (ACM template available)
   - LaTeX (preferred for technical papers)
   - Bibliography manager: BibTeX, Zotero, or Mendeley

### 4.2 Key Writing Principles

- **Lead with the problem, not the solution** — Spend first half of intro on "why this matters"
- **Show, don't tell** — Use tables, charts, examples instead of prose claims
- **Be precise about limitations** — Readers will appreciate honesty more than overselling
- **Relate to prior work explicitly** — End of each section: "Prior work X did Y; we do Z because..."
- **Make results reproducible** — Cite data availability, provide hyperparameters, release code

### 4.3 Common Pitfalls to Avoid

❌ **Writing too much about your implementation** — Focus on novel aspects, not engineering details  
❌ **Overselling results** — Be honest about what you did and didn't measure  
❌ **Ignoring limitations** — A paper that admits limitations is stronger than one that ignores them  
❌ **Poor figure quality** — Use vector graphics; label axes; provide captions  
❌ **Insufficient baselines** — Always compare to something (even simple baseline is better than none)  
❌ **Mixing methodological contribution with application** — Make clear: what's novel about the method vs. novel about the application to HVAC?

---

## Phase 5: Target Venues & Submission Strategy

### 5.1 Primary Venue: ICDAR

**International Conference on Document Analysis and Recognition**
- **Focus:** Document analysis, OCR, information extraction
- **Best fit:** Your tiered OCR pipeline + table extraction + page classification
- **Timeline:** Typically 1-2 calls per year (check ICDAR website)
- **Acceptance Rate:** ~35-40% (competitive but not impossible)
- **Papers Required:** 6-8 pages (IEEE format)

**Why ICDAR?**
- OCR + document analysis expertise in audience
- Appreciates practical applications + rigorous evaluation
- Multimodal document understanding increasingly on radar
- Open source implementations valued

**Preparation:**
- Read 3-5 recent ICDAR papers to calibrate scope & rigor
- Prepare supplementary material (code, dataset if releasable)
- Plan presentation (figures, poster format)

### 5.2 Secondary Venues

**CVPR / ICCV (Computer Vision)**
- Stronger emphasis on vision/detection aspects (floor plan YOLO component)
- Higher bar; larger pool; harder to accept
- Good for pushing detection/localization novelty

**IEEE/ACM Journals**
- Longer format (15-20 pages); deeper treatment of methodology
- Slower review cycle (3-6 months)
- More permanent venue; higher reach in industry
- Examples: IEEE Transactions on Industrial Informatics, ASCE Journal of Computing in Civil Engineering

**Construction/Facilities Management Conferences**
- ISARC (International Association for Automation and Robotics in Construction)
- Smaller audience but more domain-relevant
- Good for building on prior HVAC/BIM work (Section 8 of RESEARCH_FOUNDATION.md)

### 5.3 Submission Checklist

Before submitting:

- [ ] **Metrics:** End-to-end results clearly stated; baseline comparisons included
- [ ] **Reproducibility:** Code availability (GitHub link); dataset access (public or upon request); hyperparameters documented
- [ ] **Figures:** High-quality architecture diagram, at least one performance chart
- [ ] **Related Work:** Positioned against 10-15 most relevant papers
- [ ] **Limitations:** Explicitly discussed; not hidden
- [ ] **Author Information:** Anonymization (check venue rules; many require blind review)
- [ ] **Formatting:** Double-check venue template (margins, font, spacing, citation style)
- [ ] **References:** All citations complete; DOI/arXiv IDs provided
- [ ] **Proofs:** Spell-check, grammar check, technical accuracy review

---

## Phase 6: Post-Acceptance & Dissemination

### 6.1 Open Source Release

**GitHub Repository:**
- [ ] Publish code under permissive license (MIT or Apache 2.0)
- [ ] Write clear README (installation, usage, examples)
- [ ] Include requirements.txt with pinned versions
- [ ] Document setup (Tesseract installation, Gemini API key)
- [ ] Provide sample blueprint (anonymized / synthetic if IP-sensitive)
- [ ] Link to published paper

**Dataset Release (if possible):**
- [ ] Anonymize blueprints (remove building names, specific identifiers)
- [ ] Obtain legal review (IP, confidentiality concerns)
- [ ] Upload to Kaggle or Zenodo for discoverability
- [ ] Document labeling convention (YOLO format explanation)

### 6.2 Promotion & Impact

- [ ] Post preprint to arXiv before publication (check venue policy)
- [ ] Tweet/LinkedIn post about paper + code release
- [ ] Share with construction/facilities management communities
- [ ] Contact companies doing blueprint digitization (potential customers/partners)
- [ ] Cite in future work; encourage others to build on your approach

---

## Phase 7: Variants & Extensions (Future)

After publication, consider:

- [ ] **Journal Version:** Expand ICDAR paper to 15-20 pages for a journal
- [ ] **Dataset Paper:** Publish annotated blueprint dataset as separate contribution
- [ ] **Application Paper:** Document real-world deployment in [specific building/contractor]
- [ ] **Fine-Tuned VLM Paper:** Train custom Gemini/Llama on blueprint domain; compare to generic VLM
- [ ] **Multi-System Paper:** Generalize approach to electrical, plumbing, structural drawings

---

## Key Dates & Milestones

| Phase | Milestone | Target Date | Status |
|-------|-----------|-------------|--------|
| 1 | Research foundation complete | June 2026 ✅ | DONE |
| 2 | Paper outline complete | June 2026 ✅ | DONE |
| 3 | Experimental evaluation | July–Aug 2026 | IN PROGRESS |
| 4 | First draft (Methodology, Results) | August 2026 | PENDING |
| 4 | Full draft (all sections) | September 2026 | PENDING |
| 4 | Internal review & revision | September–October 2026 | PENDING |
| 5 | Submit to ICDAR (or target venue) | [Check deadline; typically Nov–Dec] | PENDING |
| 5 | Open source release | [After/alongside submission] | PENDING |
| 6 | Review feedback & revision | [3-4 months post-submission] | PENDING |
| 7 | Camera-ready version | [Post-acceptance] | PENDING |

---

## Resources & Tools

### Paper Writing
- **Template:** IEEE (`https://www.ieee.org/conferences/publishing/templates.html`)
- **Citation Style:** BibTeX, IEEE format
- **Collaboration:** Overleaf (free plan sufficient)
- **Version Control:** Git (track paper.tex changes)

### Experimentation
- **Tracking:** Use Excel / Notion to log experimental runs (tier performance, API costs)
- **Visualization:** Matplotlib, Seaborn, or Plotly for charts
- **Statistical Analysis:** Python scipy.stats for significance tests

### Code & Reproducibility
- **GitHub:** Public repository for source code
- **Zenodo:** Archive for versioned snapshots (one-click DOI)
- **Docker:** Consider containerization for reproducibility

### Venue Information
- **ICDAR:** https://icdar2025.org (or whatever year)
- **Papers:** Search "ICDAR" + "document analysis" or "OCR" on papers.org / arXiv

---

## Questions to Answer Before Writing

1. **How many blueprints will you evaluate?** (10 is minimum; 30+ is strong)
2. **Will you release code/data?** (Yes = more impact; No = still publishable)
3. **Do you have ground truth annotations?** (Yes = quantitative results; No = qualitative only)
4. **What's your main novel contribution?** 
   - Methodological: tiered pipeline + confidence-based fallback?
   - Domain-specific: first application to HVAC blueprints?
   - Empirical: benchmark dataset of blueprints?
   - All three? (Best position for strong paper)
5. **Who's your intended audience?**
   - Computer vision researchers?
   - Construction/facilities industry?
   - Both? (Requires different framing in different venues)

---

## Example: Publication Plan (Your Template)

**Target Venue:** ICDAR 2027  
**Paper Title:** "Automated HVAC Schedule Extraction from Mechanical Blueprints using Multi-Modal OCR and Vision-Language Models"  
**Expected Contribution:** First systematic approach + tiered pipeline methodology + benchmark dataset  
**Timeline:**  
- Experimental evaluation: July–Sept 2026
- Draft writing: Oct–Nov 2026
- Internal review: Nov–Dec 2026
- Final submission: [ICDAR 2027 deadline, typically late January]

---

## Contact & Collaboration (Optional)

If you're open to collaboration, consider:
- **Co-authors:** Domain experts in construction/facilities management?
- **Industry partners:** Contractors, architects who could provide blueprints or validate results?
- **Funding:** Research grants (NSF, DOE) for document digitization projects?

---

## Closing

You have a **strong foundation** for a publishable research paper:

✅ **Novel problem:** First systematic approach to blueprint HVAC extraction  
✅ **Sound methodology:** Tiered pipeline grounded in prior work (OCR, VLMs, object detection)  
✅ **Practical implementation:** Working code; real-world data; clear contribution  
✅ **Publishable venue:** ICDAR or construction/AI conference  

**Next immediate step:** Begin Phase 3 (experimental evaluation) and fill in PAPER_OUTLINE.md with your actual results.

Good luck! 🚀