Location: SF Bay Area / Remote

About Kedara

Millions of families struggle to coordinate elder care. We're building an AI-powered Care Coordination System with personalized, voice-AI Assistant that becomes a trusted partner for caregivers—assisting them with managing the care coordination burden and reducing burnout.

Our founders scaled Speech AI and NLP at MindMeld and BabbleLabs (both acquired by Cisco). We're a small team of product, AI/ML, and senior care experts.

The Role

We are looking for an ML engineer to own the intelligence layer across both Kedara products — Kedara Guide, our free public-facing care resource chatbot, and the Kedara App, our voice-first care coordination assistant for families.

Kedara Guide draws on public services data — local elder care resources, county programs, community services. The App merges that public context with each family's private data — care plans, medications, cultural background, language preferences. The intelligence layer handles both, and keeps them appropriately separated.

You will start by building on foundation models. Fine-tuning on care-specific and caregiver questions data is on the roadmap as we scale — including LLMs tuned to care contexts and the languages our families speak.

What you will build

LLM orchestration and RAG

• Build and maintain RAG pipelines across both products — public services context for Guide, merged public and private family context for the App
• Design the orchestration layer connecting user intent to the right AI action — retrieval, response generation, task execution, or escalation
• Integrate and manage multiple model providers — Anthropic, Google, open-source — routing based on task, latency, cost, and language
• Evaluate edge AI models vs. cloud LLMs for latency and cost-sensitive workloads
• Implement prompt versioning so changes are tracked, tested, and reversible

Stateful agent design

• Build agents that coordinate multi-step care tasks — scheduling follow-ups, syncing care plan updates, surfacing reminders at the right moment
• Design long-term memory systems so the assistant learns a family's context over time without compromising privacy
• Handle agent failure gracefully — retries, fallbacks, human escalation paths