Trafyx is what happens when a small ISP stops paying US$60,000 a year for enterprise B/OSS software written in 2011, tolerates one too many three-week support tickets, and decides to build a better one — with AI at the core, not as an afterthought.
Before we wrote a single line of Trafyx, we spent years living with the alternatives. Here's what we kept running into.
The incumbents were architected 10–15 years ago and have been getting patches ever since. Java applets, jQuery UIs, admin screens that need a training week. You can feel the age in every click. Rebuild? Not while the enterprise contracts still renew.
Ticket goes in, sits for three days, comes back with a link to a doc page that hasn't been updated since 2019. Escalate and you get someone who reads the same doc back to you. The people who actually understand the code are guarded by three tiers of triage.
Feature requests from a 400-subscriber operator sit in "backlog" forever. The same request from a carrier with 100,000 subs ships next quarter. The roadmap belongs to the biggest logo on the invoice — and if that isn't you, you get whatever's left over.
Base license, plus add-on for RADIUS, plus add-on for CPE provisioning, plus per-seat fee for each admin, plus setup fee, plus training fee. The sticker price and the actual invoice are two different products. And moving off? Data export costs extra too.
Everyone put a chat bubble on their dashboard in 2024 and called it AI. It suggests knowledge base articles. It cannot unblock a subscriber, run dunning, catch a reconciliation gap, or open a NOC ticket when a router goes dark. It's a search box in a friendly hat.
Every platform assumes one ISP, maybe a couple of sites. None of them handle the real WaaS pattern — one MSP operating shared APs with an SSID per independent ISP tenant, each with their own billing, portal, and vouchers. That gap is where we started.
Trafyx started inside Riser Technologies, an ISP running Ruckus infrastructure across hospitality, residential, and business fibre. For years we cobbled billing, RADIUS, subscriber portals, and support workflows out of six different SaaS tools that never quite fit together — while paying enterprise prices to platforms that treated us like a support ticket, not a customer.
In seventeen months we shipped a full B/OSS platform: FreeRADIUS integration, MikroTik and Ruckus provisioning, invoices and dunning, subscriber portal, vouchers, prepaid, resellers with price books, and a multi-tenant WaaS layer that routes RADIUS by SSID. Then we plugged Claude into the six places where an AI agent actually helps — ticket triage, dunning decisions, NOC monitoring, staff intelligence, portal chat, config-by-conversation — with a confidence gate and an audit note on every action.
Every feature exists because we needed it. Every safety rail exists because we tripped over its absence. The "never re-suspend an already-suspended account" rule? That's a bug we hit on day two of live dunning. The "reconciliation catcher" that flags overdue invoices which were actually paid? That's Mira noticing our own accounting drift, in preview mode, before it hit a customer.
Today Riser runs entirely on Trafyx. We use every module every day. Every feature you see was built because we needed it, tested it in production, and then decided other operators would want it too — especially the ones the incumbents forgot about.
Cron jobs, JSON contracts, whitelisted enums, dedup guards, audit logs. Nothing exciting, everything predictable. AI on top of that is safe. AI without that is a demo.
Every human minute spent on ticket triage or reading dunning reports is a minute not spent on something that grows the business. Mira handles the noise.
Every table has a tenant_id. Every query is scoped. Every UI reflects tenant. Retrofitting multi-tenancy is a nightmare — we designed for it.
"I don't know" is a better answer than a made-up one. Every Mira prompt has this line. That's why she caught a payment reconciliation bug on day one, in preview mode.
We use Anthropic's Claude for engineering, product, ops, and the Mira agent itself. It's the closest thing our team has to a force multiplier.