Streaming chat
Ask for explanations, refactors, or debug help and watch Llama 3.3 70B answer token by token over SSE.
AI code assistant · live
Codistant
Your code assistant that remembers the project.
Stream answers from Llama 3.3 70B, keep conversation memory, and run multi-step reviews - on Cloudflare, with the UI on Vercel.
What it does
Three capabilities, one assistant.
Project memory
Durable Objects with SQLite keep conversation history and project context across reloads - per session.
Multi-step review
Cloudflare Workflows run a durable three-step review: structure, issues, then concrete improvements.
How it works
A hybrid edge stack with a clear split of responsibilities.
-
01
Browser
Static UI on Vercel at codistant.v-ai.org - landing plus the chat app.
-
02
Worker
API router on Cloudflare Workers handles chat, history, context, and review.
-
03
State + orchestration
Durable Objects store memory; Workflows run the multi-step code review.
-
04
Workers AI
Llama 3.3 70B generates the streaming answers and review steps.
What I learned building this
Shipping Codistant forced real edge-AI tradeoffs - not just a chat demo.
Streaming across origins
SSE from a Worker to a Vercel origin needs CORS on every response - including the streamed body - or the browser quietly fails mid-flight.
Memory that survives reloads
Durable Objects with embedded SQLite made per-session history and project context feel local, without standing up a separate database.
Workflows for multi-step AI
Code review is three durable steps with retries. Orchestrating that as a Workflow beat stuffing everything into one fragile request.
Split hosting on purpose
UI on Vercel, intelligence on Cloudflare. Each platform does what it’s good at - and the config surface stays small.
Open to opportunities
I’m Vishal and I’m open to hire.
I build AI-native products end to end: edge runtimes, streaming UX, and practical agent workflows. If you’re hiring for AI / full-stack / platform roles, let’s talk.