Telecom Ops AI — Rob Strayer

The idea

Most AI agents sell shoes. This one runs a NOC.

The market is full of boxed agents for sales, support, and scheduling. There isn't much for the deep-technical work — the kind where a wrong answer costs an outage. The hard part isn't making an agent talk; it's making one that's right. This project's answer: keep the domain analysis fully deterministic and evidence-backed, and let the language model reason on top of it.

if X and Z then Y

A pure rules engine is trustworthy but rigid — it can't hold a conversation, weigh alternatives, or teach you the "why."

reasoning, grounded

The model reasons and converses like a senior engineer, but calls the deterministic analyzers as tools for ground truth — so it's fluent and factual. It won't invent a SIP cause or fumble subnet math.

What it feels like

A conversation, not a command line.

ops chat

calls to Verizon keep failing, here's the capture: vz.pcap
sip_analyze(vz.pcap)
cdr_analyze(today.csv, group_by=trunk)
Two things line up, and they point at the carrier, not you:
 
FACT  the egress INVITE gets a 488, and the SDP only offers G.722
FACT  trunk VZ-B answer rate is 32% today vs 85% on LUMEN-A
FACT  release cause 34 (no circuit available) dominates the failures
 
INFERENCE (high)  the far end is rejecting your codec — Verizon's
trunk isn't taking G.722. Offer G.711 as first choice, or turn on
transcoding on the SBC. The cause-34 storm says it's congestion
on their side, so open a ticket with this evidence attached.
 
Want me to generate the carrier escalation packet?

Every FACT above came from a tool that parsed the actual capture and CDR — not from the model's memory. That's the whole design: the reasoning is only as trustworthy as the evidence under it, so the evidence is computed, cited, and separated from inference.

Under the hood

Deep where it counts.

Signaling

SIP & RTP

Call ladders, failure-code interpretation, PDD & retransmission timing, two-leg B2BUA correlation, RTP loss/jitter/MOS, one-way-audio detection — from text logs or raw pcap, no tshark.

Networking

L2–L3 & systems

Route tables, virtual/VM interfaces (the bridge-MTU one-way-audio killer), SPOF analysis, BGP/OSPF adjacency diagnosis, DNS/TCP/QoS, sockets & conntrack, NTP skew.

Engineering

The bedrock math

Subnet/VLSM, Erlang-B trunk sizing, codec bandwidth & concurrent- call capacity, MTU/MSS through tunnels, TCP throughput, E.164/NANP & number portability.

Diagnostics

It goes and looks

Generates its own probes — ping, traceroute, TCP-connect, DNS/SRV, TLS certificate checks — and a fault-localizing chain that distinguishes "ICMP blocked" from "host down."

Operations

The daily grind

Work-item tracking, a deterministic daily briefing with escalation SLAs, meeting prep, alert & document intake, RCA / change-plan / capacity / failover work products with rubric linters.

Money

Chases the leaks

Circuit inventory with alias resolution, invoice reconciliation that catches zombie circuits and MRC drift, contract auto-renew sentinel, SLA credit calculation.

Principles

Private by construction.

It runs on your machine with zero dependencies — Python standard library only. Raw operational data never leaves the box; when the optional assistant layer talks to a model, phone numbers, IPs, carrier names, Call-IDs, and circuit IDs are redacted first. And it remembers — a durable memory of lessons that makes it sharper over time.

Local-first, stdlib-only, no cloud
Facts separated from inference, always
Redaction before any remote model call
Model-independent core; smart layer optional
Durable memory of operational lessons
210 tests across the analysis engine

17tagged versions

~20analysis modules

210passing tests

0dependencies

Where it's going

From tool to teammate.

The roadmap turns the sidekick into something you'd put on the team — one that not only troubleshoots but teaches, and eventually answers the phone.

Personality & a teaching mode — explains the why, from verified evidence.
Web & voice — a browser front end and WebRTC voice, so you can talk through a problem live.
Its own line — an email address and a phone number, with the telecom-compliance knowledge (A2P, STIR/SHAKEN) it already carries built in.
Self-improvement — capturing what worked, weighting lessons by outcome, sharpening its own baselines.

Who built it

Rob Strayer

Twenty-plus years of making voice and networks work when it matters — from Marine Corps voice across 160+ sites, to healthcare contact centers stood up overnight during COVID, to 911 and carrier-grade platforms, to global telecom for Microsoft's Nuance. The thread through all of it is a compulsion to understand how things really work, then build the thing that fixes them.

This agent is one of those builds — and the way he keeps learning the field: by teaching a machine to reason about it. It's the same instinct that has carried him through two decades of telecom, IT, and voice infrastructure, and into leading IT and operations at Lightspeed Voice today: figure out how something really works, then build the thing that makes it better.

LinkedIn ↗