Hi everyone,
I’m an AMD ROCm Certified Associate and independent builder (I ship projects solo alongside factory shift work). I recently finished ProjectBlock, a usage metering & Stripe billing infrastructure for AI/LLM products — lets SaaS builders track token/API usage per customer and bill automatically.
Stack: FastAPI (Fly.io), Supabase (Postgres), Upstash Redis, Next.js (Vercel), Python SDK with native LangChain callback, Stripe billing.
I’m currently listing it for sale on SideProjectors and would love the community’s honest take:
- Does this solve a real problem you’ve faced building AI products?
- Any gaps you’d flag in the approach (metering granularity, framework coverage, etc.)?
- Would this be relevant to explore with ROCm-based inference workloads down the line?
Open to any feedback — technical, market fit, or otherwise. Happy to share more details if useful.https://www.project-block.com/
Hello, on the ROCm side the gap is the cost model. Self hosted, tokens are not what you pay for, GPU hours are. A per token price does not map to anything real unless you also track utilisation. Might be worth a mode that meters tokens for billing but tracks cost per GPU hour behind it, so a builder can see the margin on each customer. I think vLLM already exposes Prometheus style metrics you could pull from, so the plumbing may not be much work.
Really appreciate the input, Hakob. You’re right that a pure token-based price doesn’t reflect what self-hosted ROCm actually costs — the GPU-hour is the real unit of cost there.
I’m going to split this out: I’ll open-source the metering piece (GPU-hour tracking wired into vLLM’s Prometheus-style metrics) as a standalone repo, and keep the billing/dashboard side of ProjectBlock closed. That way anyone self-hosting can plug in the tracking without needing the SaaS.
I’ll share the repo link here once it’s up — would be great to have you take a look or even contribute if you’re interested.
I like the idea, especially for AI SaaS products where usage-based billing can get complicated quickly. The combination of FastAPI, Redis, and Stripe seems practical, and having framework integrations could make adoption much easier. I’d be interested to see how you handle usage accuracy and edge cases like retries or failed API requests.