Project details

Tools:
Postgres, pgvector, OpenAI API, Node.js, Next.js, TypeScript, Clerk
Parentbirth: An powered AI Doula
Parentbirth started with a simple but important goal: make certified doula support accessible 24/7 without losing clinical integrity or emotional warmth.
The team had access to thousands of anonymized real-world doula conversations. These were not generic FAQs. They were nuanced, deeply personal exchanges covering pregnancy, postpartum recovery, anxiety, breastfeeding, decision-making, and emotional reassurance.
The objective was not to build a chatbot.
It was to build a trusted AI support system grounded in real expertise.
We designed a retrieval-based AI system trained on real doula transcripts, capable of delivering calm, evidence-informed, context-aware responses at any time of day. The experience needed to feel human, responsible, and safe, while operating autonomously at scale.


Challenges We Had to Overcome
Building AI in a maternal health context requires more than strong prompts. It requires architecture, constraints, and safety-first thinking.
Large language models are probabilistic. In healthcare-adjacent environments, even small hallucinations are unacceptable.
We addressed this by implementing:
Retrieval-augmented generation grounded strictly in verified doula transcripts
Postgres + pgvector vector search to anchor every response in real historical conversations
Structured prompting with explicit response constraints
Controlled temperature and deterministic fallback responses
If the system could not find grounded context, it was instructed to acknowledge limits rather than fabricate an answer

Outcomes & Wins
The result is a fully autonomous AI doula support system built on thousands of real, anonymized conversations.
Key outcomes:
24/7 availability without staffing overhead
Fully automated support grounded in real certified doula expertise
Consistent tone aligned with trained professionals
Rapid response times across pregnancy and postpartum topics
Guardrail-enforced safety boundaries