How mapped's Intelligence Works

Most AI tools bolt a single chatbot onto existing workflows and call it "AI-powered." mapped takes a fundamentally different approach: three specialized AI engines, each selected for the specific task where it delivers the highest accuracy and reliability.

The result is not a generic assistant — it's an orchestrated intelligence layer that understands methodology, respects rigor, and accelerates every step without cutting corners.

---

Multi-Engine Architecture

Every step of the systematic review workflow has different demands. A single model cannot excel at nuanced methodological judgment and high-throughput document classification and complex PDF table extraction. mapped assigns the right engine to the right task.

Advanced Reasoning Engine

Optimized for: nuanced judgment and structured analytical tasks

This engine handles the work that demands deep understanding of research methodology and careful reasoning:

• mapped VALIDATION — 8-stage, 120-point scoring against existing literature
• PICOS framework generation — structured question formulation with population, intervention, comparison, outcome, and study design
• Risk of bias assessment — consistent evaluation across RoB 2.0, ROBINS-I, NOS, and QUADAS-2 scales
• Manuscript writing — publication-ready prose that follows PRISMA reporting guidelines

Why this engine? Tasks like risk of bias demand the ability to weigh conflicting signals, maintain consistency across dozens of studies, and produce explanations that hold up to peer review. This engine was selected for its strength in structured reasoning and methodological precision.

Multimodal Extraction Engine

Optimized for: document analysis and structured data extraction from PDFs

This engine processes the visual and textual complexity of research papers:

• PDF data extraction — transforms unstructured PDFs into organized spreadsheets
• Complex table parsing — handles multi-level headers, merged cells, and nested data
• Literature search query generation — builds optimized search strings across database syntaxes
• Document analysis — extracts study characteristics, outcomes, and statistical data

Why this engine? Research papers are not plain text — they contain tables, figures, footnotes, and formatting that requires multimodal understanding. This engine was selected for its ability to see and interpret document structure, not just read words.

Classification Engine

Optimized for: high-throughput screening and rapid relevance assessment

This engine excels at fast, accurate categorization at scale:

• Title/abstract screening — rapid inclusion/exclusion decisions across thousands of records
• Full-text classification — deeper assessment of eligibility criteria
• Content generation — query optimization and search strategy refinement
• Batch processing — processes large screening queues efficiently

Why this engine? Screening often involves 5,000–50,000 records. The engine must be fast, consistent, and able to learn from your inclusion criteria. This engine was selected for its throughput and classification accuracy.

---

Why Three Engines?

A fair question. The answer is precision.

Each AI model on the market has distinct strengths: some reason more carefully, some handle documents better, some classify faster. By routing each task to the engine best suited for it, mapped delivers higher accuracy than any single-model approach.

This isn't about using more AI for the sake of it — it's about using the right AI for each step. The same principle researchers apply when choosing the right statistical test for their data.

---

Human-in-Control Principle

AI in mapped operates on a strict principle: suggest, review, confirm.

1. AI suggests — the engine generates a recommendation, draft, or classification
2. You review — every suggestion is presented for your evaluation with full transparency
3. You confirm — nothing becomes part of your research until you explicitly approve it

This applies everywhere: screening decisions, risk of bias judgments, extracted data values, manuscript text. There are no hidden AI actions.

Full Audit Trail

Every AI interaction is logged:

• What the AI suggested
• Whether you accepted, modified, or rejected it
• When the decision was made and by whom
• The final confirmed value used in your review

This audit trail ensures your systematic review meets the transparency standards required by PRISMA, Cochrane, and journal peer review. When Reviewer 2 asks "how did you arrive at this judgment?" — you have the answer.

---

Continuous Improvement

mapped's AI architecture is not static. Engines are evaluated continuously against peer-reviewed benchmarks and updated when better options become available — always prioritizing accuracy over novelty. When a model improves at a specific task, mapped adopts it. Your workflow stays the same; the intelligence behind it gets sharper.

---

Next: Learn how these engines work within each workflow step — start with Research Intelligence or see the full Workflow Overview.