Coo AI - Project Story

Inspiration

As the first parents in our friend group, we found ourselves constantly fielding questions from new and expecting parents at 3am about fever spikes, vaccine schedules, and developmental milestones, which inspired us to build an AI agent that could provide instant, evidence-based parenting guidance when pediatricians aren't available.

What it does

Coo AI is an autonomous AI agent powered by AWS Bedrock that provides personalized, 24/7 parenting guidance through SMS using intelligent 3-tier classification, age-aware symptom triage with Claude 3.5 Haiku, 5 multi-step autonomous workflows (pregnancy, vaccines, milestones, activities, preschool), and a RAG knowledge base of 70+ evidence-based medical documents from CDC, AAP, and Mayo Clinic.

How we built it

We built Coo AI on AWS using Bedrock (Claude 3.5 Haiku for reasoning, Nova Lite for classification), Lambda for serverless compute, API Gateway for routing, ChromaDB for RAG vector search, FastAPI as the web framework, SQLAlchemy for database management, and Twilio for SMS integration, with a 3-tier hybrid classification system that uses free keyword matching for 80% of queries and reserves AI for ambiguous cases to optimize costs.

Challenges we ran into

Nova Lite classification inconsistency returning explanations instead of category names (fixed with regex parsing and temperature tuning), Bedrock vs Anthropic API differences requiring unified abstraction layer, multi-turn SMS conversation state management across stateless webhooks (implemented state machine with JSON storage). Still an unsolved challenge is gathering more data.

Accomplishments that we're proud of

We built a truly autonomous agent executing complex 5-step workflows independently (not just a chatbot), achieved 99.8% cost reduction through intelligent 3-tier classification enabling us to serve 1000 families for under $3/month, implemented age-aware clinical reasoning that correctly triages fever urgency differently for 2-month-olds versus 2-year-olds, created natural language child identification understanding "my 2 year old" or "Emma" or "my baby" contextually, curated 70+ evidence-based medical documents from authoritative sources (CDC/AAP/Mayo Clinic) with no reliance on random internet scraping, achieved sub-3-second end-to-end response time from SMS to AI-generated reply despite complex processing, and deployed to production AWS in under 2 weeks.

What we learned

We discovered that hybrid AI systems combining deterministic logic with LLM reasoning outperform pure LLM solutions both in cost (500× reduction) and reliability, RAG is only as good as your data curation with 70 high-quality documents outperforming 200 random articles, system prompts for medical applications must be extremely specific about boundaries and tone to handle high-stakes scenarios safely, Nova Lite excels at structured classification tasks while Claude shines at nuanced reasoning making multi-model architecture optimal, conversation memory requires intelligent context formatting not just message storage, parents need action-oriented scannable responses under 300 characters and SMS-first design with zero friction (no app download, no account setup) is more valuable than feature-rich mobile apps.

What's next for Coo AI

Add more features to help new parents and scale the application.

Built With

• amazon
• amazon-api-gateway
• amazon-cloudwatch
• amazon-s3-|-**backend:**-python-3.11
• api-gateway-http-routing
• aws-lambda
• bedrock
• boto3
• chromadb
• claude
• fastapi
• haiku
• html/css/javascript
• iam
• mangum
• mangum-|-**database:**-sqlite-(development)
• nova
• nova-lite)
• postgresql
• postgresql-(production-ready)-|-**integrations:**-twilio-sms-api-|-**frontend:**-html/css/javascript-|-**infrastructure:**-lambda-packaging
• pydantic
• python
• sentence-transformers
• sqlalchemy
• sqlite