CogniBio

AI agent swarm replacing biopharma consulting engagements

A multi-agent reasoning framework for biopharma portfolio planning and commercial strategy. A founder types a vague KRAS question, has a brief conversation until the swarm gates open, then watches agents execute live against proprietary GTM benchmarking data — landing on a typed NPV result page with charts, sensitivities, and per-finding source attribution. Replaces traditional consulting engagements with structured, traceable, iteratively improvable analysis.

TBD

to demo

TBD

total cost

1,482

proprietary data points

Traditional vs Hot Soup

Metric	Traditional	Hot Soup
Time to insight	6-12 weeks per engagement	minutes per question
Cost per analysis	£250k+ consulting fee	marginal API cost
Auditability	PowerPoint, no trail	Langfuse span per agent call
Iteration	New SOW per refinement	Conversational follow-ups (explain, patch, reswarm)
Data integration	Manual analyst synthesis	Proprietary tool ingestion + canonical dimensions

Technical Highlights

Conversational IntakeAgent (Claude Haiku) refines vague questions through SUBJECT/STAGE/SCOPE/DECISION readiness gates
DAG pipeline: Analyst → Researcher → Synthesizer → FinancialModeler with declarative topology config
FinancialModeler produces typed KRAS NPV outputs — NPV with CI, PTRS by phase, spend curves, sensitivity factors
Per-finding source attribution (web, proprietary, database, llm) visually distinguished in the UI
Langfuse spans on every agent call for persistent observability and eval harness
Pydantic v2 schemas enforce all agent I/O contracts; Instructor adds structured-output validation + retry
185 unit tests with separate integration suite for live API calls

Tech Stack

Anthropic ClaudeInstructorFastAPINext.js 16Tailwind CSS 4shadcn/uiRechartsPydanticLangfuseRailway

Ready to build something?

Let's talk about your project and how we can help you ship faster.

Get cooking