Delfi — AI Pharma Intelligence

◈ Disease or Target

Disease or Molecular Target

Therapeutic Area

Modality Preference

◈ Candidate Intelligence

Analysing disease biology…

◈

Enter a disease to generate targets, candidates, and competitive landscape

Send to:

◎ Compound Profile

Compound / Therapy Name

Indication

Modality

Preclinical Safety Data

Clinical Data to Date

◎ Strategic Context

Competitive Landscape

Development Budget (USD)

Target Approval Timeline

Key Risk Factors

◎ Delfi R&D Scientist — Decision Report

Synthesising intelligence…

◎

Input compound profile and strategic context to receive a full GO / CONDITIONAL / NO-GO decision with toxicity, tolerability, and efficacy predictions

Actions:

Ask Delfi — Your R&D Scientist

Ask Delfi anything about drug development.

Target validation, mechanism of action, competitive landscape, trial design, regulatory strategy, toxicology, biomarkers — Delfi synthesises current pharmaceutical science to answer your questions.

💾 Saved Sessions

0 saved

💾

No saved sessions yet. After running any Candidate ID or Delfi R&D analysis, click "Save session" to preserve it here.

🔑 API Configuration

Your Anthropic API key is stored locally in your browser (localStorage) and sent only to api.anthropic.com. Get a key at console.anthropic.com.

Anthropic API Key

⚙ Fine-tuning & Examples

Provide example analyses, house style preferences, therapeutic area focus, or any guidance that should inform every Delfi response. These examples are prepended to every API request as part of the system prompt — use them to tune Delfi toward your company's voice, pipeline focus, or specific diseases you work on.

Company / Team Context

Style & Tone Preferences

Fine-tuning Examples (paste example Q&A pairs, one per line or separated by blank lines)

Preferred Therapeutic Areas (will bias analyses toward these when relevant)

ℹ About Fine-tuning

How it works: Every time you run a Delfi analysis (Candidate ID, Clinical Trials, R&D, or Ask Delfi chat), these settings are automatically prepended to the Claude API system prompt. This is called "in-context learning" or few-shot prompting — no actual model training happens, but Claude uses your examples and guidance to tailor responses.

Best practices:

Keep total context under ~2000 words — more is not always better
Provide 3–10 high-quality Q&A pairs rather than many low-quality ones
Include contrasting examples ("do this, not that")
Save often — settings persist in browser localStorage

When you have 100+ real analyses: Export them from the Sessions tab and use them as a basis for true fine-tuning via the Anthropic API's batch training endpoints, or build a RAG system on top of your exported JSON.