Autochem – Data‑Grounded Reaction Predictor

Description

Autochem is an AI-powered chemistry assistant that transforms complex reaction prediction into an intuitive, visual workflow. Users input reactants and conditions to receive comprehensive analysis including product prediction, detailed mechanisms, safety assessments, optimization recommendations, and historical benchmarking against real laboratory data. The platform combines Google Gemini's chemistry reasoning capabilities with a curated dataset of 100K+ patent reactions, delivering predictions grounded in empirical evidence rather than speculation alone.

Challenges

The primary challenge was balancing AI reasoning with empirical validation. Raw chemistry datasets required cleaning and standardization before integration. Real-time similarity matching between user predictions and patent data needed optimization for sub-second response times. Maintaining visual consistency across five interconnected workflow tabs while handling complex JSON responses from multiple AI endpoints presented significant UI/UX challenges.

Gemini Integration: Core Reasoning Engine

Autochem leverages Google Gemini 1.5 Pro as its central reasoning engine, utilizing five specialized system prompts tailored for distinct chemistry domains. The prediction prompt instructs Gemini to act as an expert synthetic chemist, returning structured JSON with product identity (SMILES notation), expected yield percentage, confidence score, reaction time estimate, and reaction classification.

Mechanism analysis employs a chemistry professor persona, generating step-by-step pathways with key intermediates and literature references. Safety assessment uses a laboratory safety specialist prompt to evaluate hazards, assign risk scores, and recommend procedures. Optimization analysis compares original versus improved conditions across yield, time, cost, and green chemistry metrics.

Gemini 1.5 Pro's 1M+ token context window enables comprehensive input synthesis (reactants, conditions, dataset context), while structured JSON output ensures reliable parsing across all endpoints. Function calling capabilities validate chemical feasibility before database persistence. Rate limiting and caching strategies prevent API overuse while maintaining responsive UX.

This integration delivers production-grade reliability, with empirical grounding from 100K+ patent reactions ensuring predictions align with real laboratory outcomes rather than isolated LLM reasoning.