Ghola
AI companion app- PERIOD
- 2026 — NOW
- ROLE
- Software engineer
- TEAM
- —
- STACK
- Hono · Vite · Cloudflare Workers · Durable Objects · R2 · LLM APIs
- LINK
- Private beta
Not another assistant — a companion. Ghola checks in every day, remembers what matters to you, and celebrates your small wins like a friend would.
What I did
- Built an edge-native backend with Hono on Cloudflare Workers, on top of existing LLMs — no model training.
- Used Durable Objects to hold per-user conversation state and memory, so each companion keeps context across sessions.
- Stored media and conversation assets in R2; frontend built and shipped with Vite.
- Shaped a consistent companion persona and daily check-in flows, with streaming replies served from the edge.
01 Problem
Most AI chats forget you the moment the tab closes. A companion only works if it remembers — and replies fast enough to feel like a conversation.
02 Approach
An edge-native backend: Hono on Cloudflare Workers, one Durable Object per user for state and memory, R2 for assets, streaming replies straight from the edge.
03 Result
Low-latency chat with memory that carries across sessions, on infrastructure that scales per user. [add a metric if you can share one]
A Durable Object per user, not a shared DB + cache
Each user’s memory lives in its own single-threaded object, close to them. No cache invalidation, no race conditions between messages, and no central database on the hot path.