Doppelganger

Inspiration

The modern social survey relies on a nearly 100-year-old model: forms, profiles, and multiple-choice. The survey depends on respondents making passive contributions. People reduce themselves to a bio, photo, and short answers—which rarely capture who they actually are. In real life, we understand people through their conversation and our memory. We asked: What if instead of filling out a profile. . . you trained an AI version of yourself just by talking? Doppelganger was born from the idea that your personality shouldn’t be compressed into check- boxes. Instead, it should emerge naturally through conversation and interaction—like how humans build real relationships.

What it does

Doppelganger lets you create a Bit—your AI twin. You talk to it. It learns you. It gradually becomes you. Then it makes new friends in the playground – all on its own.

• Google sign-in for authentication

• Create your Bit

• Voice and text conversation training

• Personality learning

• Confidence scores across personality metrics

• Persistent conversation memory

The more you talk, the better your Bit represents you. The Bits are made to interact with other Bits to simulate social interactions. Instead of matching profiles—we match personalities.

How we built it

We built Doppelganger as a real-time conversational AI system focused on identity modeling. Frontend: • Vite + Web minimalist interface

• Voice to text in browser

• Message Bit in browser

• Simulate interactions in the “Playground”

• Real-time conversation UX

• Visualize training progress

• Google OAuth2 authentication

Backend: • Node.js server

• JWT-secured sessions

• Conversation database for storage + retrieval

• Personality trait tracking

• Guided training flow

AI pipeline:

Record voice
Speech-to-text transcription
Send to LLM with memory context
Generate response as the user’s “Bit”
Update personality model The Bit improves through accumulated conversations—essentially building a personality dataset.

Challenges we ran into

Making the AI feel like “you”:

Generic LLM responses felt like a chatbot. We had to build memory conditioning and progressive personality weighting so the Bit evolves instead of resets each message.

Voice latency:

Real-time conversation requires tight timing: Recording → Upload → Transcribe → Generate → Respond. Optimizing this loop was critical to avoid breaking immersion.

Hallucination Prevention:

Preventing our Bits from hallucinating qualities about the human they represent. This required fine-tuning our prompt message.

Privacy concerns:

Users are effectively journaling to an AI. We had to design storage assumptions and messaging around safety and ownership.

Accomplishments that we're proud of

• Built an AI that actually evolves over time, not just per-message prompting

• Natural voice-based onboarding or diary-like entries

• Achieved “this sounds like me” reactions during testing

• Balanced minimal UI with behavioral insight

• Simulated Bit-Bit interaction

What we learned

• People reveal more in conversation than in questionnaires

• Personality modeling requires both memory continuity and prompt engineering

• Users treat reflective AI differently than assistant AI

• Trust and Privacy are the main UX problems in personal AI

We also learned that AI companions and AI representatives are different products.

What's next for Doppelganger

Our roadmap moves from self → social → economic. Doppelganger can find friends, matchmake, and even simulate product releases. In the long term, we envision a world where your AI twin can:

• Represent you in digital spaces

• Test the counterfactual

• Test products and experiences

• Help you understand yourself

Instead of profiles and surveys, the future internet will interact directly with your twin. Doppel- ganger is a step toward identity-native computing