Matt Siemer

GTM Engineer & Agentic AI Architect

Cincinnati, Ohio

Most GTM engineers come from engineering and learn sales. I came from sales and learned engineering.

I carried a quota for 20 years in telecom and rail. Since 2024 I’ve been building the systems I wished I’d had back then: AI agents that research accounts, clean up and enrich CRM data, and turn sales calls into follow-up tasks.

Everything I build works two ways. You can always see where a piece of data came from, and nothing gets written into a real system until a person approves it. I build it that way because a sales team won’t use a tool it doesn’t trust.

I’m looking for a full-time role in GTM engineering or forward deployed engineering, remote or in Cincinnati. You can reach me at mhsiemer@gmail.com.

25 yrsIn B2B, 20 carrying a quota

$45MRailcar fleet, full P&L

$33MTelecom equipment sold

PilotAgent running on a client’s real calls

Most recent role

Most recently I was at Clarivoy, a B2B SaaS company, from November 2024 to August 2026. I was hired to run marketing and set up the CRM. The company’s engineers were focused on its dashboard app, so nobody was building the sales and marketing systems. I taught myself to build them, and I ended up as the only builder on the company’s customer-data platform and the agent system that ran its outbound work. I pulled the CRM, call recordings, Clay enrichment and location data into one clean database with a source on every field. I built agents that answered questions about customers and reviewed contracts for the executive team, and I took a call-to-CRM sync live with zero errors on day one. Clarivoy’s data and systems are confidential, so I describe what I built, not what’s in it.

The work

Meeting Agent

Client pilot

A marketing agency records its client calls. This agent reads each one, pulls out what people committed to, and drafts those as tasks in ClickUp with an owner and a due date. A person approves every task before it posts. The team didn’t have to change anything. They file the meeting note the way they always have, and the tasks show up as drafts.

Under the hood: Python and Flask on Fly.io, Claude for extraction, token authenticated and idempotent intake so a duplicate webhook cannot double post, and a regression suite of more than 50 checks that runs the real routing rules with dry-run pushes. Writing that suite turned up four client folders that were quietly sending tasks to the wrong list. I fixed them before anything went live.

Full write-up →

The Agent Office

In daily use

This is the system I run my own business on. Dozens of AI agents work in parallel across several businesses, and each one can only do what it’s been given permission to do. Each business gets its own lane, with a written charter that says what that lane owns and where it’s allowed to make changes. One coordinating agent routes the work and keeps the records, and doesn’t build anything itself. Permission is set for each type of action, not for each agent. Some actions an agent just does and logs, some it does and tells me about, some it has to stop and ask about, and some only I can do.

Commands get checked before they run, so the rules are enforced by the system instead of only written into the instructions. Those checks have stopped my own mistakes more than once.

Full write-up →

What I know how to do

Run Clay and enrichment at scale. I’ve used Clay since September 2024. At my last company I didn’t inherit a clean database. I started from three broken, overlapping lists and built a map of the whole market from them, tens of thousands of companies, enriching and validating in Clay and keeping it current with signal monitoring. I run Clay from Claude Code instead of working row by row, and I check every result before it goes back into the CRM.
Set up a CRM and keep the data clean. I’ve built HubSpot from scratch, including custom objects and workflows, and synced it to a separate database under one rule: each piece of data has exactly one system allowed to change it. Every record shows where it came from.
Take an agent from demo to production. That means the unglamorous work that makes an agent safe to leave running: authentication, handling duplicate requests, tests that run the real rules without posting anything, a way to roll back, and an off switch.
Test the agent, not just build it. Evaluation suites, simulated users to test against, and confidence thresholds, so when the agent isn’t sure, it holds the item for a person instead of guessing.
Run a team of agents. Separate lanes, written records, different permission levels for different actions, and rules the system enforces, all documented well enough that someone else could take it over.
Work with a client to figure out what to build. Discovery, scoping, and deciding what has to be true before a build is worth doing. This is the part of the job my sales career prepared me for.
Document the work. Briefs, white papers, runbooks, and documentation that ship with the system.

Background

Systems

I build agent systems on Claude, run Clay and HubSpot through their APIs, and work in Supabase, Postgres, Python, SQL and React. Source tracking and approval steps are built in from the start.

Revenue

25 years in B2B, 20 of them carrying a quota, mostly enterprise. I ran the full P&L on a $45M railcar fleet and sold to customers including ADM, BP, AT&T, Verizon and Koch. I know pipeline mechanics, how reps actually behave, and what a long enterprise sales cycle takes.

Writing

I have a journalism degree and have written white papers, executive briefs and positioning for most of my career. I also write down why each system was built, so the next person can maintain it.

Outside work

I design and paint custom drums, including exclusive custom paint work for Ludwig. I taught myself in a condo garage and now run a custom shop.

Working stack

Languages	Python (primary), SQL, JavaScript/TypeScript, HTML/CSS
LLM and agents	Anthropic Claude API, Claude Code and Agent SDK, tool use, MCP, multi-agent orchestration, eval harnesses, prompts and rubrics kept as reviewable files
GTM stack	HubSpot (admin, API, custom objects, workflows), Clay (since September 2024: Claygents, API, credit planning), Salesforce (five years as a daily user), enrichment and validation APIs, sequencing tools
Services and APIs	Flask, REST integrations, webhooks, idempotent intake, token auth
Data	PostgreSQL, Supabase with row-level security, schema migrations, a source and date on every record
Deploy and ops	Fly.io, Netlify, Cloudflare, Docker, GitHub, secrets management, structured logging