Fairy — Hackathon Submission

Inspiration

Trying to conceive is a two-person project, but the work rarely splits evenly. Before a couple ever sees a specialist, there's a maze of insurance calls, missing lab work, and clinic scheduling, and that load usually falls on one partner. The numbers make the case: roughly 1 in 6 people worldwide experience infertility (WHO, 2023), about 1 in 5 married U.S. women trying for a first child are affected (CDC), and a single IVF cycle runs $12k–$23k, mostly out of pocket. Yet coverage is a patchwork only the patient can untangle. We wanted to give that agency back to the couple by handing the phone-and-paperwork grind to an agent.

What it does

Fairy is a mobile-first app that treats fertility prep as a shared workflow with Her, His, and Together views. From one intake, it:

• Computes a trying window and confidence label, and flags missing or borderline labs with plain-language reasons.
• Runs an autonomous agent that places a real outbound call via AgentPhone; Grok drives the conversation turn-by-turn over a webhook, then extracts a structured result (coverage, slot, bring-list) into tasks and a calendar.
• Produces a copyable, doctor-ready summary grounded only in real reference data.

The trying window uses the irregular-cycle rule from our reference data:

$$\text{ovulation}_{\min} = \text{lastPeriod} + \text{cycleMin} - 14$$

$$\text{ovulation}_{\max} = \text{lastPeriod} + \text{cycleMax} - 14$$

$$\text{window} = [\ \text{ovMin} - 5,\ \text{ovMax} + 1\ ]$$

How we built it

Two people on parallel branches in one day, with Cursor in the loop throughout.

• AgentPhone for real outbound voice; xAI Grok as the conversational brain on each webhook turn, plus reasoning, structured extraction, and grounded chat.
• Inngest for a durable, event-driven workflow: parallel fan-out/fan-in, a waitForEvent approval gate, a step.sleep check-in, and a reactive summary function, all driven by four events.
• Call-time RAG: xAI embeddings + a Redis vector KNN retrieve the exact reference knowledge (CPT codes, coverage terms) into the agent's prompt, off the Supabase hot path.
• Supabase for data, Vercel for deploy, Next.js App Router + Tailwind for the phone-frame UI.

A core design rule: a pure rules core (trying window, missing-data, extractors) with no I/O, wrapped by an impure shell, so correctness is testable in isolation.

Challenges we ran into

• Live voice is messy. Humans answer out of order, vaguely, or not at all. We treated the call as a checklist of objectives rather than a script, and parse the transcript regardless of order, marking unresolved fields instead of inventing them.
• Demos can't stall. Any live failure (no key, bad webhook, dropped call) falls through to a deterministic Mock_Fallback that returns identical schema and values, so the demo always completes.
• No fabricated medicine. Every clinical value comes from /reference-data/; if it isn't sourced, it isn't shown.
• Keeping two parallel branches coherent under a tight clock without stepping on each other's seams.

Accomplishments that we're proud of

• A voice agent that makes a real phone call and writes back structured, grounded results.
• An Inngest graph that is visibly more than a one-shot function: parallel branches, a human-in-the-loop pause, and a scheduled follow-up.
• 201 tests with the rules core covered by property-based tests (any-order extraction, idempotent approval, parallel fan-in), tsc clean and build green.
• A polished, native-feeling phone UI, not a desktop dashboard.

What we learned

• Determinism is a feature, not a fallback. A reproducible mock made the live path safe to demo.
• Agentic ≠ linear. Modeling the flow as events with pauses and parallelism matched the real-world process far better than a sequential pipeline.
• Grounding every claim in source data is what makes health AI trustworthy, and it's a constraint worth designing around from the start.

What's next for Fairy

• Real telephony and identity verification (with a signed BAA and HIPAA compliance) to move beyond demo numbers.
• Multi-couple support with secure, per-partner data ownership.
• Deeper personalization from labs, cycle history, and biometrics, plus expanded coverage-comparison across plans and clinics.

Built With

• cursor
• grok
• inngest
• vercel
• xai