Hi, I’m
Spencer.
I build products
with and for AI.

I’m a product manager in New York City. After years working on portfolio risk modeling software, I’ve spent the last year turning AI ideas into working software.

Now: exploring new product roles in NYC, or remote.

Let’s go blueNYC

Spencer Rosenberg wearing a blue baseball cap

Product manager. AI builder. New York City.

Resume

Open PDF Download LinkedIn

For the past year

Founder and product builder

Built a few different projects ranging from AI observability, agentic web improvements, and personalized outreach.

January 2020 to July 2025

Morgan Stanley

I mostly worked as a product manager on portfolio risk management software for our wealth advisors and E*TRADE clients. Left voluntarily to pursue AI startups.

Education

University of Michigan

BA Economics, class of 2019.

Why should you hire me? Ask my agent.

Spencer’s agent

Ask why Spencer could be a good fit for your team.

My startup experiments,
a retrospective.

Teckel

Could LLMs train smaller models to monitor enterprise AI?

What I tried: Users could prompt what they wanted to track in their deployed AI systems, have an LLM find traces, and train a small custom model to cheaply detect behaviors such as prompt injection or PII in responses. I conceived the product idea early on when using LLMs as judges for observability was considered expensive and inefficient.

What happened: AI observability became crowded, while enterprise demand shifted toward end-to-end agent platforms. Most companies did not want to build agents internally and add a separate monitoring SDK. Teckel was also designed before tool calls and long reasoning traces became standard. The original implementation could not manage this added monitoring complexity well.

camolabs

Could a website tell visiting agents what they should do?

What I tried: I built a system that let websites detect, monitor, and route agent visits, then publish actions agents could call, such as booking a demo or requesting a proposal from the team. It was designed as an alternative to poorly utilized standards such as WebMCP and llms.txt files.

What happened: Getting agents to do what you want is structurally inconsistent. Action discovery remained unreliable, even with dedicated pages served through Cloudflare edge routing. Agents often ignored unfamiliar routes, were wary of steering that could be similar to prompt injection, rarely identified themselves correctly when visiting a website, and often fetched cached pages with stale information.

Ask your agent to run one of these MCP actions on this website

POST /mcp

Open actions directory

01
call_tool get_resume
Returns my resume as structured data.
02
call_tool email_spencer { sender_name, sender_email, … }
Send me a job opportunity or introduction.
03
call_tool check_open_times { date?, time_zone? }
Returns my 30-minute Cal.com availability.
04
call_tool book_meeting { slot_id, name, email, … }
Book a selected time on my calendar.

Camo

Could we turn a job post into a personalized invitation?

What I tried: Camo was built for recruiters. It copies a company’s visual design language, logo, and style, and builds a one-off invitation page with a clear call to action for a specific role.

What happened: It works, but I do not think it is a venture-scale opportunity. You can try the product here.

Stuff I've accomplished as a product manager.

At Morgan Stanley, I worked on portfolio risk modeling software for our industry-leading wealth management business. My stakeholders included financial advisors, high-touch clients, engineers, risk partners, sales teams, and senior management.

$4.5Tassets under management modeled daily by our platform

$192Mnet new advisory assets in one quarter, attributed to our platform

25.2%year-over-year usage increase after I started leading the growth team

90%core stakeholder adoption firm wide