Skip to content

PROCESS.md

Latest commit

 

History

History
86 lines (65 loc) · 3.28 KB

PROCESS.md

File metadata and controls

86 lines (65 loc) · 3.28 KB

🚀 Development Process - From Problem to Production

Week 1: Problem Discovery & Validation

Goal: Validate the pain point is real and measurable

Activities:

  • Interviewed 3 NYC general contractors
  • Discovered "insurance renewal fire drills" pattern
  • Measured: 40-80 hours per project searching OpenSpace captures
  • Validated willingness to pay: $500-1000/month for solution

Key Insight: The problem wasn't classification accuracy, but evidence organization for auditors.

Week 2: MVP Definition & Architecture

Goal: Build smallest possible solution that delivers value

Technical Choices:

  1. Frontend: Streamlit (over React/Django) - fastest to ship
  2. AI/ML: DeepSeek (over OpenAI) - 90% cost savings
  3. Deployment: Streamlit Cloud - zero DevOps overhead
  4. Data: NYC Open Data API - free, authoritative source

MVP Scope: Upload → Classify → Generate Report (no bells/whistles)

Week 3: Core Development

Technical Challenges & Solutions:

Challenge 1: Noisy Construction Images

Problem: Dust, poor lighting, motion blur reduced classification accuracy to 60%

Solution:

  • Implemented image preprocessing (brightness/contrast normalization)
  • Added confidence scoring with fallback logic
  • Created domain-specific prompt engineering

Result: 85%+ accuracy on real construction photos

Challenge 2: NYC-Specific Compliance Rules

Problem: Generic construction classification missed NYC DOB requirements

Solution:

  • Built NYC DOB code reference database
  • Added compliance gap detection algorithm
  • Integrated live violation data from NYC Open Data

Challenge 3: Cost Management

Problem: OpenAI GPT-4 would cost $5-10 per project (prohibitive)

Solution:

  • Switched to DeepSeek (90% cheaper)
  • Implemented caching for repeat analyses
  • Added batch processing to reduce API calls

Week 4: Polish & Presentation

Focus: Making it portfolio-ready

  1. Professional UI: Custom CSS, Plotly charts, responsive design
  2. Demo Data: Sample project folder for instant testing
  3. Documentation: README, PROCESS.md, TECHNICAL_ARCHITECTURE.md
  4. Deployment: Live at https://sentinelscope.streamlit.app/

What I Learned

Technical Insights:

  1. LLMs are feature extractors, not problem solvers - you need business logic on top
  2. Prompt engineering is iterative - version control your prompts
  3. Mock data enables demos - crucial for portfolio projects
  4. Streamlit is amazing for MVPs - but has scaling limits

Product Insights:

  1. Solve the user's problem, not the technical challenge
  2. Measure everything - hours saved, dollars saved, accuracy rates
  3. Ship fast, gather feedback, iterate
  4. A mediocre solution today beats a perfect solution never

What I'd Do Differently

  1. Start with user testing earlier - built some features that weren't critical
  2. Implement error tracking from day 1 - lost some debugging time
  3. Write more tests earlier - technical debt accumulated
  4. Document as I code - spent time reconstructing decisions

Next Steps

  1. User testing with 2-3 contractors for feedback
  2. Batch processing for large projects (1000+ images)
  3. OpenSpace API integration (vs CSV upload)
  4. Multi-project dashboard for GCs with multiple sites