AI Research System Architecture & Enhancement Strategy
A comprehensive analysis of our current multi-agent VC research system built on Google's Gemini LLM with LangChain orchestration, along with strategic recommendations to dramatically improve accuracy, reliability, and user experience through advanced prompting techniques and verification layers.
Current System Overview
Multi-Agent Architecture
Specialized research agents working in parallel, each focused on distinct aspects: people, product, company, and competitive landscape
Dual LLM Strategy
Gemini 2.5 Pro for deep research requiring accuracy, Gemini 2.5 Flash for faster validation and creative synthesis tasks
Intelligent Orchestration
LangChain-powered tool-calling agents that autonomously decide when to search, scrape, and synthesize information
The system processes company research requests through a sophisticated pipeline that validates inputs, conducts parallel deep research across multiple dimensions, and synthesizes findings into comprehensive investment memos with actionable feedback.
Research Pipeline Architecture
01
Preliminary Validation
Gemini Flash validates company existence and founder links using 10 web searches. Requires 80% confidence to proceed, preventing wasted resources on invalid targets.
02
Parallel Deep Research
Four specialized agents execute simultaneously: People Research (15 searches), Product Analysis (13 searches), Company Background (17 searches), Competitive Analysis (15 searches).
03
Synthesis & Feedback
Gemini Flash generates 3-5 critical discussion points after deduplicating sources and analyzing all reports for gaps, contradictions, and hidden risks.
Core Intelligence System
Weekly Portfolio Monitoring
An automated intelligence pipeline that tracks portfolio companies through multiple data sources and generates prioritized recommendations for VC partners.
  • LinkedIn company activity scraping (7-day window)
  • Market intelligence scouting with recency filters
  • Automated priority assessment and action recommendations
  • Curated podcast suggestions for market context
Current System Strengths
1
Parallel Execution Design
ThreadPoolExecutor runs four research tasks simultaneously, dramatically reducing total processing time while maintaining research depth and quality.
2
Structured Output Validation
Pydantic schemas with OutputFixingParser ensure valid JSON responses. Auto-correction with 2 retry attempts prevents malformed data from breaking downstream processes.
3
Intelligent Search Budgets
Carefully calibrated iteration limits control costs while ensuring sufficient research depth. Different budgets per research type optimize the accuracy-cost tradeoff.
4
Confidence-Based Gating
Preliminary validation prevents expensive downstream processing on invalid companies. Confidence scores guide resource allocation throughout the pipeline.
Critical Enhancement: Few-Shot Learning
1
Current State: Instruction-Only
Models receive detailed instructions but no concrete examples, leading to inconsistent output quality and missed nuances in desired format and depth.
2
Enhanced Approach: 2-3 Full Examples
Include complete, high-quality examples in each major prompt. Models learn style, depth, and format far more effectively from examples than instructions alone.
3
Advanced: User-Provided Examples
Allow VCs to submit their own exemplar reports. Build modular specialist agents that parent agents intelligently select based on the specific research needs.
Impact: Few-shot learning dramatically improves output quality, consistency, and alignment with user expectations. The token cost is absolutely justified by the performance gains across all research modules.
Verification Architecture: Multi-Layer Fact Checking
Current Limitations
Simple prompt instructions like "No fabrication" are ineffective guardrails. Models may still hallucinate facts, especially under pressure to fill required fields or meet confidence thresholds.

Key Issue: Telling models to avoid hallucination can paradoxically increase hallucination by priming the concept.
Enhanced Verification
  1. URL-Fact Verification: Second-pass LLM confirms each fact matches its cited source (parallelized for speed)
  1. Quote Attribution: Require specific quotes as backing, enforce structural text matching against source material
  1. Cross-Reference Validation: Multiple sources required for critical claims, automated consistency checking
Agent-Critic Architecture
Primary Agent
Performs tool calls, conducts research, reasons through information gathering
Critic Review
Every N messages, equal or higher intelligence LLM reviews progress and research paths
Steering Decision
Critic identifies dead ends, suggests course corrections, validates research direction
Final Criticism
Synthesized report undergoes adversarial review: attack LLM vs defense LLM with judge
This approach prevents agents from pursuing unproductive paths while maintaining research momentum. The scrolling window with progressive summaries keeps costs manageable while preserving context.
User Experience Enhancements
Natural Language Refinement
Users express preferences or complaints in natural language. LLM reformats feedback into system prompt adjustments for next generation, eliminating rigid output dissatisfaction.
Intelligent News Deduplication
Weekly reports compare against previous N reports using embedding similarity or targeted LLM comparison. System emphasizes that 0-2 news items is acceptable, preventing forced content generation.
Remove Preliminary Validation
Trust user input by default. If company isn't found, surface this as a research finding rather than blocking execution. Eliminates unnecessary friction and false negatives.
Implementation Roadmap
1
Phase 1: Quick Wins
  • Add 2-3 few-shot examples to each prompt
  • Remove preliminary validation gate
  • Implement news deduplication
Timeline: 2-3 weeks
2
Phase 2: Verification Layer
  • Build URL-fact verification system
  • Add quote attribution requirements
  • Implement parallelized fact checking
Timeline: 4-6 weeks
3
Phase 3: Advanced Architecture
  • Deploy agent-critic framework
  • Build adversarial review system
  • Add natural language refinement
Timeline: 6-8 weeks
4
Phase 4: Customization
  • User-provided example system
  • Modular specialist agents
  • Intelligent agent selection
Timeline: 8-12 weeks
This phased approach allows for incremental improvements while maintaining system stability. Each phase delivers measurable accuracy and reliability gains that justify continued investment in the enhancement roadmap.