Orbit — an AI that learns who you actually are

The problem

Every social app matches you on what you claim to be — bios, hobby checklists, your best photo. But people describe their ideal self, not their actual self. So matches are made between two fictional people, and that's why they die after "hey."

"Social apps reward people who present well.
Orbit works for people who don't."

How Orbit works

Interviews you — seven conversational questions about what you did, not what you're like. "Remember the last time your plans got cancelled — what did you actually do, and how did you feel?"
Shows its reasoning — a live panel of every inference, each backed by your own words as evidence, each with an honest confidence level. One trait is never asked at all: communication depth is measured from how you answered.
Takes correction — anything wrong, one click fixes it. The user is the final authority on themselves. We call it Proof of Personality.
Matches you — hard filters (intent, dealbreakers — never overridden), embedding scoring, then Gemma names the connection type: mirror, complement, growth, opposite-world, shared journey. Never a percentage — a person, a reason, a first activity.
Introduces you — on mutual acceptance, Orbit writes the introduction itself. The AI as mutual friend. No more empty "hey."

On AMD hardware

0.8 s

per synthetic user profile
(13 LLM calls each)

2.1 min

to interview & analyse
148 synthetic users

148×

concurrency headroom
(162 GiB KV cache)

Privacy by hardware, not by policy. The interview collects intimate personal detail. Everything — interview, extraction, matching, introductions — runs on a single MI300X via vLLM/ROCm. No external LLM API at any point; the data never leaves the box. Same pipeline on a consumer laptop: ~16 minutes per profile.

Run the live app yourself

Orbit's interactive demo runs against a live LLM (the MI300X droplet, or Ollama locally), so this page is a static overview — the full application is three commands away:

git clone https://github.com/Logicrithm/orbit-ai
pip install -r requirements.txt
uvicorn app:app        # open http://localhost:8000

Point .env at any OpenAI-compatible endpoint — MI300X + vLLM (see setup.md, rebuilds in ~10 minutes) or local Ollama with gemma3:4b.

Honest limits

Synthetic population

148 Gemma-role-played personas test the pipeline, not real-world match quality. Simulated elements are labeled on screen.

Heuristic weights

The staged funnel exists so weights can be tuned against real feedback — which is designed, not yet built.

A better prior

Behavioural questions are harder to fake than trait claims — not impossible. Orbit is a better prior, not a truth machine.