/ en / work / cost-observability
Cost Observability — this site, live
How I run agentic AI on a $5/day budget — with a public kill switch
→ Business impact: AI budget controlled to the cent: per-visitor / per-session / site caps, hard kill switch, public costs on /eval.
- Client
- nyami.fr — this site
- Role
- Architect & operator
- Period
- 2026
- Status
- live
- $5/day site budget
- $0.03/day visitor cap
- ~$0.0003 cost/answer
- ARMED kill switch
Business context
A public agent wired to a paid LLM is an open credit card on the internet. Here, cost control is the product — running live on this site.
Constraints
- Structurally capped budget
- Zero spend without a verified quota
- Public transparency
- Free edge worker, no dedicated database
Architecture
Architecture diagram — anonymised
Components: Cloudflare Workers · KV · TypeScript · SSE
Control workflow
Every request passes the caps (visitor, session, site) BEFORE the LLM call; every run is accounted; counters feed the public dashboard and the kill switch.
Key architecture decisions
- Quotas consumed before spend — a provider failure never refunds the slot
- Multi-level kill switch (auto day/month + manual)
- Transparency by default: spend and state are public
My contribution
- Full architecture and implementation
- Per-visitor-tier cost policy
- Public dashboard and no-auth endpoint
Outcomes
- Structural $5/day cap
- Cost per answer (~$0.0003) shown to the visitor
- Verifiable by anyone, live
Confidentiality
Everything is public — /eval and the cost endpoint are open.
Learnings
- A cost constraint becomes a product argument when it is visible.
- Pre-spend guardrails are the only guaranteed ones.
🚧 V-next — already in the works
This system ships continuously — the next stages are already on the bench. Come back to check the status, or ask the agent where it stands.
- Multi-model routing — planned for V1.1.
- Multi-threshold email alerts — activation under way.
- Hardened real-time counter aggregation — V1.1.