// bioinformatics ai

FailedExp Agent

A sophisticated bioinformatics AI system for root-cause analysis of failed laboratory experiments. Transforms raw experiment descriptions into actionable troubleshooting recommendations using LLM reasoning, knowledge retrieval, and pattern detection.

try now →

System Architecture

FailedExp employs a multi-stage agent pipeline powered by LLM reasoning, knowledge retrieval, and pattern detection. The system orchestrates six sequential nodes, each transforming experiment data into structured insights.

Technology Stack

→ Frontend: Next.js 14 + TypeScript + React
→ Backend: FastAPI + Python 3.11 async/await
→ Database: PostgreSQL + SQLAlchemy ORM
→ Vector Store: ChromaDB for semantic search
→ LLM: Claude 3.5 Sonnet API
→ Workflow: LangGraph for deterministic pipelines
→ Deployment: Render + Vercel

Pipeline Overview

1. Parse — Unstructured text → Structured metadata
2. Retrieve — Augment with peer-reviewed protocols (ChromaDB RAG)
3. Detect Patterns — Taxonomy-based classification (11+ experiment types)
4. Hypothesize — LLM chain-of-thought reasoning with evidence citations
5. Recommend — Actionable recommendations prioritized by confidence
6. Store — Persist analysis to PostgreSQL for audit trail

Key Features

Explainability

Every hypothesis has evidence citations. Users see exactly why the system reached its conclusions, critical for laboratory decision-making.

Reproducibility

Deterministic rule-based patterns ensure consistent results. Same input → same patterns (no ML variance).

Extensibility

New techniques added as rules + templates. Current: 11 experiment types. Future-proof architecture for 20+ types.

Scale

Async Python + PostgreSQL + ChromaDB supports 1000+ concurrent requests. Sub-20s latency on standard servers.

Performance

End-to-end analysis pipeline with detailed timing breakdown:

Stage	Time	Bottleneck
Parse	2-3s	LLM API latency
Retrieve	1-2s	ChromaDB vector search
Detect	0.5s	Rule matching (deterministic)
Hypothesize	4-6s	LLM reasoning
Recommend	2-3s	LLM personalization
Total	10-15s	LLM inference (60%)

Technical Highlights

Smart Temperature Parsing

Pydantic validator automatically parses temperature strings in any format: "95°C" → 95.0, "98 °C" → 98.0, "95" → 95.0. Handles user input variations gracefully.

Control Detection Logic

Missing or failed controls are RED FLAGS. Absence of negative controls alone suggests contamination risk and is weighted heavily in the failure pattern taxonomy.

Retrieval-Augmented Generation (RAG)

Instead of relying solely on Claude's training data, augments with curated peer-reviewed protocols and institutional troubleshooting guides via ChromaDB semantic search.

Ready to analyze your failed experiments?

open failedexp →