PersonaShift

💡 Inspiration Most AI tools today try to tell human leaders what to do—generating confident, single-answer recommendations that flatten nuance and amplify AI sycophancy. In high-stakes organizational, civic, and policy dilemmas, there is rarely a single "correct" answer. Every major choice involves trade-offs that benefit some parties while burdening others.

We built PersonaShift to solve this problem. Instead of automating away human judgment, PersonaShift acts as a decision-analysis workstation that maps the human and operational landscape of a dilemma before a commitment is made.

⚙️ What It Does PersonaShift takes an ambiguous dilemma (e.g., "Our college is considering making attendance mandatory" or "The city is converting a major street to pedestrian-only") and guides the user through a disciplined 4-stage workflow:

MAP (Stakeholder Engine & Human-in-the-Loop Review): Discovers affected parties categorized into Direct, Indirect, and Systemic stakeholders. The human retains full authority to keep, edit, remove, or add custom stakeholders before proceeding. SHIFT (Signature Lens Switching & Epistemic Tracking): Generates structured multi-dimensional perspectives (Goals, Concerns, Constraints, Incentives, Priorities). Every single claim is explicitly labeled with its epistemic certainty: [FACT] — Explicitly stated in the source problem. [INFERENCE] — Logical contextual deduction. [UNKNOWN] — Recognized information gap that cannot be asserted as fact. The signature interaction: Users can switch between stakeholder viewpoints with 0ms latency and zero API calls, instantly seeing the problem through each party's eyes. HEAR PERSPECTIVE (Multimodal deAPI Audio): Users can audibly absorb perspectives using deAPI's text-to-speech API. It delivers an objective, neutral third-person audio synthesis ("Possible factors shaping the Students perspective...")—avoiding cartoonish first-person roleplay. COMPARE (Systemic Tensions & Dependencies): Surfaces cross-cutting dynamics, including shared goals, differing priorities, a dyadic conflict matrix of potential tensions, and operational dependencies. EXPLORE (Grounded Alternatives without AI Bias): Generates distinct policy proposals (PROPOSAL I, II, III). Every addressed concern is mathematically grounded in verified stakeholder perspectives, presenting realistic trade-offs and implementation considerations without prescribing a single "winner".

🛠️ How We Built It Frontend: React 18, Vite, Vanilla CSS (custom high-contrast design system, fully responsive across 390px–1440px, WCAG AA accessible). Backend: Node.js (ES Modules), Express REST API. Reasoning Engine: Google Gemini (gemini-2.5-flash / @google/genai SDK) utilizing strict JSON schemas for structured outputs. Multimodal Audio: deAPI Text-to-Speech API (https://api.deapi.ai/api/v2/tts) integrated via an Express proxy with strict host whitelisting and SSRF protection. Data Integrity & Validation: Strict runtime validation using Zod across all API boundaries and LLM responses. Testing: 7 comprehensive automated test suites covering parser extraction, stakeholder taxonomy, perspective invariants, zero-latency SHIFT switching, comparison logic, concern grounding, and deAPI audio security.

🧗 Challenges We Ran Into Preventing LLM Hallucinations in Alternatives: In the Exploration stage, models often invent convenient compromises that don't match stakeholder needs. We implemented strict concern grounding algorithms that reject proposals citing non-existent or cross-attributed concerns. Eliminating Sycophancy & Roleplay: Standard LLMs tend to roleplay ("As a student, I feel...") or give generic advice. We enforced epistemic discipline (FACT vs. INFERENCE vs. UNKNOWN) and objective third-person framing throughout prompts and narration scripts. Audio Streaming & SSRF Security: Proxying audio streams from deAPI storage to circumvent browser CORS required building a hardened proxy (/api/perspective-audio/proxy) that strictly validates HTTPS requests against results.deapi.ai, rejecting internal IPs, loopback, or spoofed domains.

🏆 Accomplishments That We're Proud Of The SHIFT Primitive: Perspective switching happens in-memory with zero network overhead, making perspective-taking feel like an intuitive software tool rather than waiting on another chatbot prompt. Anti-AI-Slop Aesthetics: Delivered an authoritative, dark-slate editorial workstation aesthetic (inspired by tools like Linear and Palantir) that avoids generic gradients, neon colors, glassmorphism, or AI sparkles. 100% Invariant Verification: Passed all 7 automated test suites with live API and security regression tests.

📚 What We Learned How to combine Google Gemini's structured reasoning with deAPI's multimodal voice synthesis to create a unified decision-support workflow. Why epistemic humility ([UNKNOWN]) is essential for building AI systems that human decision-makers can actually trust.

🔮 What's Next for PersonaShift Executive Dossier Export: One-click export of structured decision briefs in PDF/Markdown format for board meetings. Multi-User Stakeholder Alignment: Allowing real stakeholders to input their actual concerns and comparing them against AI-modeled inferences. Policy Longitudinal Tracking: Tracking decisions over time to compare real-world outcomes against modeled trade-offs.