Where it all starts
First Input, getting the facts together
Second Input, subjective
Third Input, emotions
Fourth Input, expectations

FiAu — Fair Intake & Alignment Understanding

Inspiration

The Problem

Every couple has that one recurring argument they’ve had forty times. Whether it starts over dirty dishes, unmade plans, or a late text, surface triggers quickly spiral into defensive walls.

Mid-fight:

Defensiveness takes over: Both partners attempt to "win" the argument rather than understand it.
Venting loops repeat: Raw emotional statements escalate tension and widen the gap.
Miscommunication dominates: What is intended as a request is heard as an attack.

The Lightbulb Moment

Standard AI chatbots often fail in relationship mediation because they take sides, hallucinate verdicts, or leak harsh phrasing. We built FiAu (Fair Intake & Alignment Understanding) to answer a single question:

"What if AI wasn't a judge or a therapist, but a neutral, empathetic mirror that maps where two people actually align and diverge—without ever taking a side?"

🚀 What It Does

FiAu transforms chaotic relationship friction into structured, productive resolution through a 5-Step Guided Workflow:

Private 24-Hour Ephemeral Rooms: One partner generates a temporary 6-character room code. No accounts, no emails, and zero long-term data tracking. Rooms auto-delete after 24 hours.
4-Layer Psychological Intake: Each partner reflects independently across four structured layers (What happened, What did you feel, What did you make it mean, and What do you want) using keyboard input or our hands-free Hold-to-Talk voice instrument.
Sealed Words Privacy Model: Neither partner ever sees the other’s raw written text. Raw words are sealed locally and processed exclusively by the AI engine. Partners only view synthesized, non-defensive reframes, completely eliminating knee-jerk defensive reactions.
Crisis Safety Interstitial Layer: Automatic keyword scanner checks intake text for harm or crisis language, immediately pausing AI submission to display calm, direct crisis support helplines.
The 4-Act Narrative Resolution Scroll:
- Act 1: Surface Proxy Card: Reframes surface triggers ("dirty dishes") into root emotional drivers ("feeling unappreciated during stressful work weeks").
- Act 2: Agreements Bridge: Displays a sage-green baseline of shared ground and verbatim facts where partners agree before addressing any differences. (Explicitly labeled as Skipped in Solo mode).
- Act 3: Interactive Structure Map & Presentation Views: Dissects divergent statements across an interactive ReactFlow node graph (IssueTree), a sticky side-panel for verbatim quote verification and absolute wording reframes ("always" → "often"), a Fullscreen Overlay Canvas, and a step-by-step Slide Deck Presentation View.
- Act 4: Handback Closing Letter: Formats an actionable closing letter with recommended discussion steps ("Facts first, values last") and a 1-click copy button for reading aloud together.

How We Built It

Frontend & UX: Next.js 16 (App Router), React 19, TypeScript, and Tailwind CSS v4. Interactive node mapping is powered by @xyflow/react and framer-motion.
Voice Capture Instrument: Designed a custom HoldToTalk voice recorder utilizing the Web Speech API and Web Audio API (AnalyserNode frequency spectrum) for real-time live waveform animation with zero audio persistence.
Storage Seam: Built an ephemeral Store interface with Upstash Redis REST for production and an automatic InMemoryStore fallback for local dev resilience.
Multi-Model AI Pipeline: Structured a 3-stage sequential LLM pipeline:
- Call 1 (Extract): Extracts structured claims per person with verbatim quote enforcement (Primary: Mistral AI ministral-8b-latest, Fallback: Groq openai/gpt-oss-120b).
- Call 2 (Align): Discovers perspective alignments and surface proxies.
- Call 3 (Handback): Generates closing scripts and action sequences under strict adjudication checks.

Challenges We Ran Into

Enforcing Strict AI Neutrality: Prompt engineering alone wasn't enough to guarantee neutrality. We built automated regex scanners (findAdjudication) that scan LLM output for verdict-y words ("right", "wrong", "at fault"). If detected, the pipeline automatically re-runs Call 3 or applies forceNeutralize() to sanitize output.
Eliminating AI Hallucinations (Verbatim Provenance): We wrote strict provenance invariants (enforceProvenance) that verify every extracted quote against original intake text. Any claim whose quote is not an exact verbatim substring is automatically dropped.
Widescreen & Laptop Canvas Optimization: Interactive node graphs can feel cramped on small screens or sparse on large laptop displays. We solved this by implementing responsive 12-column desktop split grids, a Fullscreen Overlay Mode, and a Slide Deck Presentation View.

Accomplishments That We're Proud Of

Zero-Account Privacy Architecture: Native 24-hour TTL lifecycle where data auto-deletes from memory.
Sealed Words Innovation: Proved that hiding raw emotional text and presenting only synthesized reframes lowers partner defensiveness immediately.
Fluid Dual Presentation: Created two complementary ways to explore Act 3—an interactive node graph for visual thinkers and a sequential slide deck for linear readers.
100% Clean TypeScript Infrastructure: Achieved 0 type errors across complex multi-stage AI pipelines and fallback stores.

What We Learned

UI Boundaries Create Safety: Psychological safety isn't just about soft copy—it's enforced by UI constraints (sealing raw words, capping voice recordings to 90 seconds, and limiting takes).
Multi-Stage LLM Pipelines Outperform Single Prompts: Splitting AI tasks into dedicated Extract, Align, and Handback calls with strict Zod validation produces vastly higher quality and more empathetic outcomes than single-prompt wrappers.

What's Next for FiAu — Fair Intake & Alignment Understanding

Multilingual Cross-Cultural Support: Real-time translation so partners speaking different native languages can reflect in their own language and receive a unified map.
Audio Narrative Readout: AI voice synthesis to read the Act 4 closing letter aloud for guided joint listening sessions.
Counselor & Therapist Export: Standardized PDF summary export for couples who want to bring their resolution map into professional therapy sessions.