The solution enables faster, guideline-compliant protocol creation, boosting researcher productivity and accelerating time-to-market for new therapies.
Meet our client
Client:
Industry:
Market:
Technology:
In a Nutshell
Client’s Challenge
The client needed to automate the creation of highly regulated Study Protocols, a process that traditionally takes months of manual drafting. The main challenge was generating accurate, large-scale medical text while ensuring strict compliance with ENCePP and internal regulatory guidelines.
Our Solution
We developed an LLM-powered application that automates guideline-compliant Study Protocol creation, integrating directly with the client’s internal standards. The system combines LLMs models with RAG from PubMed, Semantic Scholar, ClinicalTrials.gov, and internal databases to ensure accuracy and relevance. Built with Python, FastAPI, PostgreSQL, and S3, the MVP is now ready for testing and accelerates protocol development from months to weeks.
Client’s Benefits
The solution enables faster, guideline-compliant protocol creation, boosting researcher productivity and accelerating time-to-market for new therapies.
A Deep Dive
1) Overview
Project name: ENCePP-Aligned Protocol Generation
Short description:
Powered by an LLM, the system generates complete ENCePP-aligned protocol drafts, reducing authoring timelines from months to weeks.
To strengthen scientific validity, the system incorporates retrieval-augmented generation (RAG) drawing on PubMed, Semantic Scholar, ClinicalTrials.gov, and internal research databases to surface contextually relevant evidence and ensure methodological accuracy.
A production, guideline-aware authoring system that generates complete study protocol drafts aligned with internal guidance and ENCePP-style structure—used by client teams to compress timelines from months to weeks.
Key objectives
- Generate complete protocol drafts across core sections (Objectives, Study Design, Population, Data Sources, Outcomes, Analysis, Ethics).
- Encode internal guidance so drafts follow required structure and phrasing from the outset.
- Fit seamlessly into existing authoring habits and exports.
Key outcomes
- In production and used by clients today.
- Cycle-time reduction: months → weeks.
- Specialists spend less time on administrative drafting and more on scientific work.
2) Client
A global pharmaceutical R&D organization operating across multiple therapeutic areas. The teams run observational and post-marketing studies at scale and emphasize consistent document quality, guideline compliance, and rapid study start-up.
3) Challenge
Ensuring evidence-grounded drafting required integrating external biomedical knowledge sources, necessitating a RAG pipeline capable of retrieving peer-reviewed literature, prior study designs, and registry information from PubMed, Semantic Scholar, ClinicalTrials.gov, and internal datasets.
Business challenge
Manual protocol drafting demands months of specialized effort and coordination. Shortening this process accelerates evidence generation and downstream study execution.
Technology challenge
Automatically producing high-quality medical-domain text at volume while consistently adhering to internal SOPs and ENCePP-aligned structure—without disrupting established authoring practices.
4) Solution
The system uses the LLM frontier model to generate guideline-aware protocol drafts aligned with ENCePP structure and internal SOPs. LLM controllability and ability to follow complex templates ensure high first-draft completeness and consistent terminology.
Protocol generation and refinement
- Template-driven inputs: Users select protocol templates and provide key parameters (e.g., study design, endpoints, population, data sources).
- Guideline-aware drafting: The system generates each section to match internal guidance and ENCePP-style structure, ensuring required elements are present and phrasing is consistent.
- Evidence retrieval via RAG: During drafting, the system can surface relevant publications, prior protocols, and registry entries from PubMed, Semantic Scholar, ClinicalTrials.gov, and internal sources to support justification of design choices and improve scientific completeness.
- Section cohesion: Cross-references (objectives ↔ outcomes ↔ analysis) are harmonized to maintain internal consistency across the document.
- Review & improvement: Once the draft is generated, users can:
- Edit sections manually in the editor, or
- Use an AI assistant to refine, expand, or rephrase individual sections with targeted prompts.
- Exports: One-click export to client-standard DOCX/PDF for distribution and archival.
Selected technologies
Python, FastAPI, Pydantic, PostgreSQL, Amazon S3, LLM-based text generation and reviewing.
Solution diagram
5) Process
We collaborated with teams to encode internal guidance into structured prompts designed for LLMs, enabling the model to reliably follow ENCePP-style methodological standards and produce sections with consistent structure and phrasing.
- Discovery, problem understanding & scoping – Map internal guidelines, templates, and success criteria.
- Prompt engineering from internal templates & guidelines – Encode structure, tone, and required elements into reusable prompts and templates.
- Integrate retrieval workflows: Configure RAG pipelines to retrieve high-quality external evidence from PubMed, Semantic Scholar, ClinicalTrials.gov, and internal repositories, enabling the model to incorporate contextual information into draft generation when requested.
- Create an MVP to gather initial feedback – Release to early users, collect edits and usability input to calibrate generation quality.
- Productionization & ongoing support – Harden for scale, integrate exports, and provide continuous updates based on real-world usage.
6) Outcome
LLMs ability to maintain internal document cohesion (e.g., linking objectives, endpoints, and analysis plans) reduced review cycles and increased confidence in methodological consistency.
Quantitative (observed/expected)
- Drafting lead time reduced from months → weeks for first-pass documents.
- Fewer review iterations due to standardized structure and terminology.
Qualitative
- Teams focus more on scientific reasoning and less on administrative writing.
- Higher confidence that drafts align with internal guidance and ENCePP-style expectations from the start.
- Evidence-grounded drafting improved justification quality, as RAG-sourced references and examples reduced back-and-forth cycles related to missing context or insufficient methodological rationale.
Lessons learned
- Encoding guidance inside templates and prompts yields higher first-draft completeness than post-hoc fixes.
- Keeping the authoring loop close to writers (manual edits + AI assist) drives adoption and quality.
7) Strategic Impact & Next Steps
Future iterations will leverage agentic capabilities to perform self-review against rubric-based criteria, propose targeted revisions, and retrieve relevant prior protocols to improve consistency.
Strategic impact
- Faster evidence generation: Shorter drafting cycles bring study starts forward and reduce coordination overhead.
- Consistency by design: Templates and guidance encode structure and phrasing, improving baseline quality and reducing rework.
- Institutionalized know-how: Reusable prompts and templates capture internal best practices and scale them across teams and TAs.
- Scalable operations: A single system supports diverse protocol types while keeping outputs coherent and compliant.
- Commercial advantage: More predictable timelines and higher first-draft quality enhance bid competitiveness and partner confidence.
Next steps
- Improved validation
- Deeper structural checks for required elements and cross-references.
- Terminology libraries and phrasing standardization.
- Integrity checks for assumptions, endpoints, and analysis plans, with clear “trust-but-verify” reports.
- More autonomous / self-improving / agentic approach
- Iterative draft refinement loops that propose targeted edits.
- Self-evaluation against rubric-based criteria, feeding back into prompts and templates.
- Retrieval of relevant prior protocols and reusable fragments to improve coherence and speed.
- Team collaboration
- Shared workspaces and draft sharing to streamline collective input.
- Lightweight tasking and activity views to coordinate who improves which sections.
8) Summary
A production, guideline-aware protocol generator that creates complete, consistent drafts and streamlines refinement through manual edits or AI assistance, helping client teams move from drafting to study execution significantly faster.
