Skip to content

VENTURELENS_JOURNAL_EN.md

Latest commit

 

History

History
199 lines (139 loc) · 3.15 KB

VENTURELENS_JOURNAL_EN.md

File metadata and controls

199 lines (139 loc) · 3.15 KB
Raw

🚀 VentureLens Project Journal (English)

1. Initial Idea (VentureLens-1)

The original goal was simple:

Build an AI tool that can analyze startup ideas.

At this stage, the system relied mostly on:

  • prompting an LLM
  • generating general startup advice

❌ Problem

The outputs looked good but had a major flaw:

  • no grounding in real data
  • answers were generic and sometimes unrealistic

💡 Insight

“If the system only uses LLM, it’s just a chatbot — not an analysis system.”

2. Transition to Data-Driven Approach (VentureLens-2)

Introduced real startup dataset:

  • description
  • industry
  • funding
  • success_score
  • outcome_label

Key shift

From: Idea → LLM → Output

To: Idea → Data → Signals → Output

3. Retrieval System (VentureLens-3)

Implemented TF-IDF similarity.

What worked:

  • system started referencing real startups

❌ Problems:

  • noisy matches
  • dataset bugs (empty columns, wrong formats)
  • still relied on LLM

Fix:

  • data cleaning pipeline
  • standardized schema

4. Full Pipeline Thinking (VentureLens-4)

Redesigned system as a pipeline:

Idea
→ Retrieval
→ Scoring
→ Risk
→ Scenario
→ Recommendation
→ Report

Insight:

Separate:

  • data (retrieval)
  • logic (rules)
  • generation (language)

5. RAG System (VentureLens-5)

Added rag_context_builder.py

Responsibilities:

  • retrieve similar startups
  • compute signals:
    • avg_success_score
    • success_ratio
    • failure_ratio
    • similarity

Key idea:

LLM is no longer the brain, only the narrator.

6. Scoring System Evolution

❌ Problem:

Random ideas got very high scores.

Root cause:

  • weak penalties
  • no caps
  • overly optimistic aggregation

✅ Fix:

  • penalty system
  • similarity thresholds
  • score caps

Insight:

Scoring must be conservative, not optimistic.

7. Risk Analyzer

Rule-based instead of LLM:

Signals:

  • failure_ratio
  • similarity strength
  • peer count

Outputs:

  • risk_level
  • risk_flags

8. Scenario Simulation

Upgraded from simple text → structured output:

Each scenario includes:

  • title
  • description
  • why it happens
  • trigger
  • warning signs
  • actions

9. UI (Streamlit)

Components:

  • score + radar chart
  • peer signals
  • similar startups
  • scenarios
  • recommendations

Insight:

UI must explain reasoning, not just output.

10. Final Architecture

ai_modules/

  • data_loader.py
  • similarity_bridge.py
  • rag_context_builder.py
  • scoring_engine.py
  • risk_analyzer.py
  • scenario_simulator.py
  • recommendation_engine.py
  • report_generator.py
  • full_analysis.py
  • radar_chart.py

11. Key Lessons

  1. LLM without data = unreliable
  2. Retrieval quality is critical
  3. Scoring must be calibrated
  4. Structure > complexity
  5. Most bugs come from data mismatch

12. Future Improvements

  • Embeddings instead of TF-IDF
  • Industry baseline scoring
  • Percentile ranking
  • Better evaluation
  • Memory system

🎯 Interview Summary

“I built a startup analysis system that moved from LLM-based guessing to a data-driven RAG pipeline with scoring, risk modeling, and structured scenario simulation.”