Last week I saw a post from
@lablabai promoting a hackathon using
@AIatAMD @Alibaba_Qwen tools… and it instantly sparked an idea: build a real
#personalAIagent.
I spent the next 5 days planning it and 2 develolpin. I got inspired by
@0xJuliechen 's project — a chat that felt like you were actually talking to a famous person through their quotes. That magic made me create @fiwing_chat .
Here’s the architecture I’m building:
The inbound path is powered by Spectrum Cloud stream consumption via
@photon_hq 🔥 This decision removes one external UI dependency from the critical request path.
User identity is normalized early:
@supabase user id(UUID) is the canonical key. Phone = routing. space_id = conversation thread context.
Clear separation. No duplicate users. No cross-thread contamination. No identity headaches.
Memory path is fully per-user and pulled on every single turn thanks to
@honchodotdev. Context retrieval isn’t “best effort” — it’s an explicit, non-negotiable stage before generation.
Inference runtime runs on
@FireworksAI_HQ through
@AMD using the
@Alibaba_Qwen family (model selection via env vars). Switching providers or models is now just a config change, not code surgery.
@composio is mandatory in the execution flow. If the required tools aren’t available, the system degrades explicitly instead of hallucinating capabilities.
Security boundary: the app only stores ca_* references. OAuth tokens stay inside Composio’s boundary — minimizing blast radius.
Operationally, this is deploy-safe, observable, and built with explicit control points.
Day 1 delivered a solid systems baseline: channel identity memory model tooling, all integrated.
This is how you build a true Digital Twin that actually feels personal and secure.
Who else is building their own AI agent right now? Drop your stack below 👇
#BuildInPublic #AI #DigitalTwin #AMD #PHOTONHQ