Penngate Deep Research System
Multi-agent coordination producing comprehensive, citation-backed research documents of 12,000+ words with 50-100 verified references—in approximately 3 hours instead of 50-80 hours of human work.
Three-Tier Agent Architecture
Structured autonomy with clear role separation and vertical accountability
Planner Tier
1 Agent
The Research Planner decomposes topics into 3-7 domains, creates specialist briefs with research questions and suggested Perplexity queries.
Specialist Tier
5 Agents × N Domains
- Literature Scout
- Data Analyst
- Domain Expert
- Synthesis Writer
- Fact Checker
Evaluation Tier
3 Agents
Entity Extractor builds the knowledge graph. Quality Judge evaluates and triggers retries. Report Compiler assembles the final 12,000+ word document.
The Nine Agents
Each agent has a narrow focus and operates with clear acceptance criteria
Research Planner
Decomposes topics into 3-7 domains, creates specialist briefs with research questions and suggested queries.
Literature Scout
Searches Perplexity for academic sources, scrapes via Firecrawl, prioritises high-reliability sources.
Data Analyst
Extracts quantitative data and statistics, generates charts, identifies benchmarks and trends.
Domain Expert
Provides deep contextual analysis, maps historical context, identifies key debates and controversies.
Synthesis Writer
Transforms findings into coherent prose with natural flow, maintains citations and source attribution.
Fact Checker
Verifies claims against cited sources, detects conflicts between specialists, resolves contradictions.
Entity Extractor
Extracts named entities, infers relationships, stores globally in Neo4j with provenance tracking.
Quality Judge
Evaluates quality on 5 dimensions, triggers targeted retries for weak domains only (max 10 cycles).
Report Compiler
4-phase document assembly: Preparation, Writing, Assembly, Polish. Produces 12,000+ word output.
AI Research Pipeline
Multi-agent system workflow orchestrated by LangGraph
Research Planner orchestrates 5 parallel specialists per domain, with entity extraction and quality evaluation before final compilation
Fan-Out / Fan-In Architecture
5N parallel tasks with self-correction loop
Enterprise-Grade Infrastructure
Built for 3-4 hour research tasks with fault tolerance and observability
LangGraph + LangSmith
Orchestration & Observability
StateGraph with Send API for fan-out/fan-in parallelism.
LangSmith provides token tracking, trace collection, and executive insights.
Neo4j + PostgreSQL
Knowledge Graph & State
Neo4j models entities, relationships, and citation networks with provenance tracking.
PostgreSQL provides checkpoint recovery and state persistence across failures.
Celery + Redis
Task Queue & Events
Handles 3-4 hour research tasks with 5N parallel specialist operations.
Real-time progress streaming via SSE with fault-tolerant message delivery.
Human vs AI Research
Phase-by-phase breakdown of research workflow
| Phase | Human Researcher | Penngate System | Reduction |
|---|---|---|---|
| Topic Scoping & Research Questions | 4-8 hours | 15-30 minutes | ~93% |
| Literature Search & Source Discovery | 8-15 hours | 10-20 minutes | ~97% |
| Reading, Annotation & Analysis | 12-20 hours | 30-60 minutes | ~95% |
| Synthesis & Interpretation | 6-12 hours | 20-40 minutes | ~94% |
| Outlining & Structure | 1-2 hours | 5-10 minutes | ~90% |
| Writing First Draft | 6-10 hours | 20-40 minutes | ~94% |
| Revision, Editing & Fact-Checking | 5-10 hours | 15-30 minutes | ~95% |
| Final Compilation & Formatting | 2-3 hours | 5-10 minutes | ~93% |
| Total Focused Work | 50-80 hours | ~2-3.5 hours | ~95% |
| Calendar Time | 8-16 weeks | Same day | ~99% |
Concurrency & Self-Correction
Two levels of parallelism compound to deliver massive speed-ups
Intra-Cycle Parallelism
5 specialists operate concurrently per domain using LangGraph's Send API for fan-out.
Cross-Domain Parallelism
Tasks across all N domains dispatched simultaneously—5N total parallel operations.
Self-Correction Loop
Quality Judge triggers targeted retries for weak domains only—up to 10 cycles.
Real-World Use Cases
How organisations leverage multi-agent research across industries
Healthcare & Clinical Research
Systematic literature reviews for clinical evidence, drug efficacy meta-analyses, treatment protocol research, and epidemiological studies.
How It Works
- Planner decomposes into domains: clinical trials, drug interactions, patient outcomes
- Literature Scout searches PubMed, Cochrane Library, and clinical databases
- Data Analyst extracts statistical evidence and outcome measures
- Fact Checker validates all claims against source material
Key Benefits
- Accelerates systematic reviews from months to hours
- Cross-references multiple clinical databases automatically
- Audit-ready citation trails for regulatory compliance
- Supports evidence-based decision making
Defence & Government Intelligence
Threat assessment research, defence technology reviews, policy analysis, safety case evidence compilation (Def Stan 00-056).
How It Works
- Domain Expert provides classified-context analysis
- Literature Scout scans defence journals and government publications
- Quality Judge enforces security classification compliance
- Report Compiler produces standards-compliant documentation
Key Benefits
- Rapid threat landscape synthesis
- Safety case evidence compilation
- Standards compliance (Def Stan, MIL-STD)
- Intelligence briefing acceleration
Financial Markets & Investment
Quarterly market analysis, sector deep-dives, ESG research, risk assessment reports, M&A due diligence research.
How It Works
- Data Analyst extracts market metrics and benchmarks
- Literature Scout searches financial databases and SEC filings
- Synthesis Writer produces investment-grade prose with citations
- Report Compiler formats for client-ready delivery
Key Benefits
- Reduces analyst report time from 2 weeks to hours
- Consistent cross-quarter formatting
- Automatic cross-referencing of data sources
- Audit trail compliance for regulated environments
Management Consulting & Strategy
Market entry analysis, competitive landscape mapping, industry trend reports, technology due diligence.
How It Works
- Research Planner identifies market segments and competitive dimensions
- Domain Expert provides industry context and analysis frameworks
- Literature Scout discovers market reports and competitor data
- Report Compiler produces client-ready deliverables
Key Benefits
- Client-ready research packages in hours
- Multi-source market intelligence
- Standardised report formatting
- Supports proposal development and thought leadership
Observability & Monitoring
Full visibility into agent performance and costs
Time Reduction
AI Research Time
Words per Document
Verified Citations
Delivery
Ready to Transform Your Research Workflows?
The Penngate Deep Research System delivers comprehensive, citation-backed research documents in hours instead of months. Built for government, healthcare, financial services, and consulting organisations.