What I Wanted To Build
The idea behind AI Exodus is one primitive: a model of one specific person that is good enough to speak for them. Dating is the first product because the stakes are easy to understand. If a twin can represent you well enough to find out in two minutes that you and someone else have nothing to say to each other, nobody has to waste a Friday evening finding out in person.
That only works if the twin is faithful and the matching is honest. Most of the engineering in this project goes into those two things.
How The App Is Put Together
- Client: a React 18 app built with Vite and wrapped in Capacitor for Android. The same code serves the web app. Training scenes are played by a small visual-novel player built on framer-motion.
- Backend: Python 3.10 Lambda functions deployed with the Serverless Framework in eu-west-3 (Paris), behind API Gateway, with a single DynamoDB table and Cognito for accounts.
- Models: Gemini Flash for the twins' conversations and for grading them, and Gemini embeddings (768 dimensions) for the twin's memory.
- Background work: EventBridge schedules for the nightly matching sweep, and asynchronously invoked workers for the long-running twin conversations.
Training A Twin Without A Questionnaire
Filling in a form produces the person you want to be. A scene produces the person you are. The app ships ten branching story scenes (ten stories with 167 scene nodes in total, from a short "first date" intro to a long career crossroads) where you pick what you would actually do. The player records your choices and syncs them to your profile, and there is a separate chat-style trainer where you correct the twin when it speaks wrong.
What the twin learns is stored as memories: rows with an embedding, ranked on retrieval by similarity, importance and recency. Corrections and highlights become memories too, so telling the twin "I would never say that" actually sticks. When a twin speaks, it pulls the twelve most relevant memories rather than a keyword match, which was one of the shortcomings of the first version I rewrote.
Letting Twins Talk First
A "hunt" is a background job. The request returns immediately (API Gateway gives a function 29 seconds, and a twin conversation takes longer), and a worker with a 5-minute limit does the real work. It first filters candidates with cheap, free, structured checks: shared likes, clashes between likes and dislikes, and a minimum amount of twin training. Only candidates that survive are worth a model call, and a conversation costs about six cents. Then the two twins talk for a bounded number of turns, each guided by a director prompt that steers the phase of the conversation. Private data never enters the twin-to-twin prompts, and the twins talk in English whatever language their owners use, with the transcript translated for the human reading it.
The Bugs That Taught Me Most
- Every conversation scored exactly 50. The judge that grades a twin conversation was a model with "thinking" enabled, and the thinking used up the whole output budget. The reply came back empty, the JSON failed to parse, and a hardcoded fallback of 50 was returned for every hunt. For a while I even lowered the match threshold to compensate, which made matches worse. The fix was to turn thinking off for the judge, ask for JSON directly and keep the reply short. Afterwards a good conversation scored 95, a lukewarm one 85 and a dead one 15.
- Politeness scored as chemistry. Two twins are always nice to each other, so a plain "be stricter" instruction drifted. I made the judge list the topics that actually landed and the ones that were declined, then enforced a hard rule in code: if no topic landed, the score is capped at 35 whatever the model said.
- "Best of one" is random. A hunt used to take the first candidate. Now it runs up to three conversations and surfaces the best, which only became worthwhile once the grader worked. Three conversations at about 45 seconds each fit inside the worker's 300-second limit.
- "Unknown" scored as a perfect match. Missing attachment data was counting as the best possible pairing. The structured score now leaves out dimensions where either profile has no data and reports a confidence value instead.
- Cost control. Coins are charged only when a match is found: 1 for a manual hunt and 5 for an auto-hunt, with a cheap structured gate before any model call and a cap on how many hunts the nightly sweep starts. A cheap gate was calibrated on 516 real candidates, where the median raw compatibility was 0% and only 3.7% reached 10%.
Where It Is Honestly Unfinished
- It is in Early Access. I do not yet have measured data on how many twin matches turn into real conversations between people, so I make no claim about that.
- The conversation design defines phases up to forty turns, but conversations are capped far lower, so the later "values" and "emotional" phases are mostly unreachable today.
- Some pieces are placeholders, such as the readiness check for the vouch feature, and some memory features, like capturing a memory from every single chat message, are still on the list.
- Twin-to-twin text is not moderated beyond the model's own safety settings, so how much to trust a conversation is still partly the human's job.
- Plans for a larger architecture (queues, real-time sockets) exist as design documents only. What runs today is the Lambda and DynamoDB stack described above.
If you want to try it, the app is free to start. If you want to talk about twin-style personalisation, agentic conversations on a budget or shipping a Capacitor app, use the quote form on the home page or ask the assistant there.
▶ Get AI Exodus on Google Play →