Conversation Lab

Inspiration

How do dating apps match people? Despite their sophisticated algorithms, most apps just end up showing you a never-ending loop of people who look like your previous "swipes." It’s a surface-level experience that trains users to value appearance over personality and connection.

What if it was more? What if we matched people based on their IDENTITY. We believe compatibility is found in the way people talk, their unique humour, and their core personality.

By analyzing the "Conversation DNA" of a user, we allow people to discover matches they are actually likely to enjoy hanging out with. No more wasting time, effort, and money on endless back-and-forth messages that lead to failed dates. We help you find the connection before the first "Hello."

What it does

Conversation Lab analyzes chat history from dating apps to extract a user's "Conversation DNA" - their unique communication fingerprint. This includes 20+ traits across 6 dimensions: style (emoji use, formality), interaction patterns (question rate, response depth), social signals (warmth, humour), topics of interest, conflict handling style, and emotional needs.

When comparing two people, we calculate compatibility scores across these dimensions and generate AI-powered insights explaining why a match might work and what to watch out for. Users of the website can also simulate a hypothetical first conversation between two people based on their DNA profiles, however this would only be used as a stress-test to catch contradictions and help produce more vivid insight in the backend and would not be shown to users of the app.

How we built it

We built the frontend with Next.js and React, styled with Tailwind CSS. The AI pipeline uses LangGraph to orchestrate multi-step analysis:

1. Parse & validate chat history
2. Extract Conversation DNA using GPT-4o (structured JSON output)
3. Calculate compatibility scores using deterministic math for reproducibility
4. Generate human-readable insights with GPT-4o
5. Optionally simulate conversations using Claude Sonnet

We used Backboard.io as our LLM gateway, which let us seamlessly switch between OpenAI and Anthropic models without managing multiple API keys.

Challenges we ran into

The biggest challenge was making the AI output consistent and structured. Early versions would return malformed JSON or inconsistent DNA profiles. We overcame this by implementing strict Zod schemas, rigorous temperature tuning, and fallback handling to ensure the data remained reliable.

We also navigated complex ethical and legal considerations. We decided on a strictly opt-in only model where users must explicitly consent to their data being used, alongside a "right to be forgotten" feature where users can delete their data at any time.

We also hit a LangGraph "UNREACHABLE_NODE" error when trying to make the simulation step conditional. We fixed this by splitting the pipeline into two separate graphs instead of using dynamic routing.

Accomplishments that we're proud of

• Created a hybrid scoring system that combines deterministic math (for more accurate reproducibility) with AI insights (for human-readable explanations)
• Built realistic conversation simulations that actually reflect each person's communication style
• Designed a clean, intuitive UI that makes complex AI analysis feel approachable
• The whole pipeline runs in a reasonable time despite making 3-4 LLM calls

What we learned

• LangGraph is powerful for orchestrating multi-step AI workflows, but requires careful state management
• Structured output prompting is an art - being too specific can make the AI rigid, too vague makes it inconsistent
• Backboard.io made it easy to experiment with different models without code changes

What's next for Conversation Lab

• Real-World Integration: Partnering with dating app APIs to analyze real-time conversation flows (only upon user opt-in).
• The "Vibe" Algorithm: Revolutionizing discovery by showing users people whose core identities naturally resonate with theirs first.
• Outcome Fine-Tuning: Training our models on actual relationship outcome data to move from "predicted" to "proven" compatibility.
• Deepening the Lab: Adding new dimensions to our profiles, including Attachment Styles and Love Languages.
• Expansion: Taking the Lab beyond dating into professional networking and high-performance team building.
• Identity-Stress Testing: Running 2-3 simultaneous simulations to catch contradictions and ensure our analysis reflects a user's true persona.
• Continuous Improvement: Ongoing UI/UX evolution to keep the "Conversation Lab" experience at the cutting edge.

Built With

• anthropic-claude
• backboard.io
• langgraph
• next.js
• openai-gpt-4o
• react
• tailwind-css
• typescript