MIRA TRL Assessment Pipeline

Automated Technology Readiness Level (TRL) assessment for biotech projects on the Molecule platform.

MIRA fetches documents from a project's dataroom, uses AI to extract and analyze their content, classifies the project into a TRL stage, scores it against weighted criteria, and generates actionable improvement todos.

How it works

The pipeline processes each project through five sequential AI steps:

Dataroom Documents (PDFs, images)
        |
        v
  1. Content Extraction         (Claude Sonnet 4.6)
        |
        v
  2. Questionnaire Filling      (Claude Opus 4.6)
        |
        v
  3. TRL Classification         (Claude Opus 4.6)
        |
        v
  4. Weighted Scoring            (Claude Opus 4.6)
        |
        v
  5. Todo Generation             (Claude Opus 4.6)

Step 1: Content Extraction

Each public document (PDF, PNG, JPEG, GIF, WebP) is downloaded and sent to Claude Sonnet 4.6 with a prompt focused on extracting scientifically relevant content. The model returns structured output containing:

• Document type (pitch deck, experiment report, publication, etc.)
• Relevance to TRL assessment (high / medium / low / none)
• Content blocks organized by page, each typed as text, figure, table, or diagram
• Key findings for TRL assessment

Extractions are cached to disk as JSON files keyed by a SHA-256 hash of the file's identity (did:version:path). On subsequent runs, unchanged files are loaded from cache instead of re-extracted.

All extracted documents are then merged into a single formatted text that preserves file boundaries, page structure, and block types. Documents with relevance: "none" are excluded.

Step 2: Questionnaire Filling

The merged content is sent to Claude Opus 4.6 with a biotech project questionnaire covering:

• Project overview -- title, therapeutic concept, target indication, patient population, mechanism of action, biological target, development timeline
• Experimental evidence -- lab work status, experiment types, results, model systems (in vitro, advanced in vitro, in vivo), efficacy evidence, dose-response data, reproducibility, safety observations
• Team information -- key personnel, affiliations, expertise
• Intellectual property -- patent status, scope of protection, freedom-to-operate

The model fills each field based on what it found in the documents. Fields that cannot be determined are marked "Not specified".

Step 3: TRL Classification

The completed questionnaire is sent to Claude Opus 4.6 with a systematic decision tree that classifies the project into one of five levels:

Level	Name	Definition
Pre-TRL 1	No coherent hypothesis	Project lacks a coherent scientific hypothesis
TRL 1	Concept Alone	Purely theoretical -- no experimental evidence exists
TRL 2	Speculative Research	Some experiments conducted, but only in simplified systems (e.g., basic in vitro)
TRL 3	Early Evidence	Proof-of-concept demonstrated in disease-relevant models (in vivo or advanced in vitro)
TRL >3	Beyond scope	IND-enabling studies, GMP manufacturing established

The decision tree follows three questions:

1. Has any experimental lab work been conducted? No -> TRL 1
2. Have experiments been conducted in disease-relevant in vivo models or advanced disease-modeling systems? No -> TRL 2
3. Do results demonstrate early proof-of-concept efficacy in disease models? No -> TRL 2, Yes -> TRL 3

When ambiguous, the lower TRL is assigned (conservative principle).

The output includes: trl_classification, confidence (high/medium/low), rationale, and key_indicators.

Step 4: Weighted Scoring

Only projects classified as TRL 1, 2, or 3 are scored (Pre-TRL 1 and TRL >3 are skipped).

Each TRL level has a different set of criteria with different weights. The raw project data is sent to Claude Opus 4.6 along with the TRL-specific scoring rubric.

TRL 1 (3 criteria):

Criterion	Weight
Therapeutic Relevance of Mechanism of Action	60%
Therapeutic Optionality	15%
Intellectual Property	25%

TRL 2 (5 criteria):

Criterion	Weight
Therapeutic Relevance of Mechanism of Action	45%
Therapeutic Optionality	5%
Intellectual Property	15%
Utility of Candidates	20%
Prospects for Safety	15%

TRL 3 (9 criteria):

Criterion	Weight
Therapeutic Relevance of Mechanism of Action	20%
Therapeutic Optionality	5%
Intellectual Property	15%
Utility of Candidates	20%
Prospects for Safety	15%
Prospects for GMP/CMC for IND Filing	5%
Prospects for Clinical Development	10%
Commercial Potential	5%
Organization and Team Fit	5%

Each criterion is scored 1-5 (5 = strongest), multiplied by its weight, and summed into a final_weighted_score (max 5.0). The model also produces a score_interpretation summary.

Step 5: Todo Generation

For any criterion that scored below 5, the pipeline generates actionable improvement items. The model receives the full scoring context plus the raw dataroom content and produces 2-20 todos.

Each todo contains:

Field	Description
criterion_key	The scoring criterion this addresses
title	Short actionable title (max 80 chars)
description	2-3 sentences referencing specific dataroom files
effort	low (upload existing), medium (compile/format), high (new experiments/filings)
specificity	Exactly what document or evidence to upload
type	public (non-sensitive, AI-reviewable) or private (confidential, team-verified only)

Key generation rules:

• Effort is evidence-based: "low" only if the item is explicitly referenced in the dataroom. If there's no evidence it exists, it's "high".
• Impact maps to criterion weight: >=15% = high impact, 10-14% = medium, <10% = low.
• Scope-aware: Todos respect the project's stated roadmap and current development phase. A TRL 1 project won't get todos about GMP manufacturing.
• Ordered by actionability: Immediately uploadable items first, complex new work last.

Project structure

src/index.ts          Main entry point and pipeline orchestrator
lib/config.ts         Constants: model IDs, API endpoints, contract addresses, test projects
lib/llm.ts            All AI logic: extraction, questionnaire, classification, scoring, todos
lib/molecule.ts       Molecule GraphQL API client (fetches projects and dataroom files)
lib/sanity.ts         Sanity CMS integration (storage and change detection)
output/               Cached extraction results, organized by project ID

Tech stack

Dependency	Purpose
Vercel AI SDK (ai)	Unified interface for LLM calls with structured output via Zod schemas
@ai-sdk/anthropic	Anthropic provider -- Claude Sonnet 4.6 for extraction, Claude Opus 4.6 for reasoning
Zod v4	Schema validation for all structured LLM outputs
TypeScript	Strict mode, ES2021, NodeNext modules
Node.js 22	Runtime
pnpm	Package manager

Running locally

pnpm install

Create a .env file:

ANTHROPIC_API_KEY=...
MOLECULE_API_KEY=...
SANITY_API_TOKEN=...

Run against the hardcoded test projects:

pnpm dev

Run against all projects from the Molecule API:

pnpm dev:all

Environment variables

Variable	Required	Description
ANTHROPIC_API_KEY	Yes	Anthropic API key for Claude models
MOLECULE_API_KEY	Yes	API key for Molecule's GraphQL endpoint
SANITY_API_TOKEN	Yes	Authentication token for Sanity CMS
DISCORD_WEBHOOK_URL	No	Discord webhook for notifications (production only)
SANITY_STUDIO_URL	No	Base URL for Sanity Studio (required for Discord notification links)
ENVIRONMENT	No	staging or production -- controls Sanity dataset
USE_ALL_PROJECTS	No	true to fetch all projects; otherwise uses test projects
FORCE_PROCESS	No	true to skip change detection and reprocess everything

Scheduled execution

A GitHub Actions workflow (.github/workflows/trl-assessment.yml) runs the pipeline daily at 08:00 UTC. It processes all projects, commits updated extraction caches back to the repo, and can also be triggered manually with configurable options for environment, project scope, and force-processing.