Skip to content
workingmemory

Resume

Rob Belcher

Austin, TX · rbelcher@gmail.com · linkedin.com/in/rob-belcher · workingmemory.yarbros.net/work

Senior / Principal Product Manager · AI agents on live PayPal calls · Several production AI apps I built and run

Product manager who ships LLM agents, writes the code, and designs whole systems rather than features. At PayPal I'm spearheading the real-time AI agent layer on live customer calls for a global 9,000-person contact center serving 400M+ customers: the prompts, the KPIs, and the evals. On my own I build and run several AI apps in daily use, from a meal-planning and grocery-shopping agent to a job-search system, with live production figures at workingmemory.yarbros.net/dashboard.

Independent Projects

2026 · live production figures at workingmemory.yarbros.net/dashboard

Several AI apps in daily production for my family and outside testers, built end-to-end with Claude Code on shared sign-in, email, data, and CI services I built underneath them. Every agent is isolated, privacy-tested on each deploy, and instrumented for turns, tokens, and cost.

  • YarMeals plans a family of five's week and shops two grocery stores unattended, asking only what it can't already answer.

    Ended weeks of fast food for a family of five. It plans the week from a rated recipe catalog, matches every line to a real product at two stores with no API, and loads both carts through a browser-driving agent that a scripted orchestrator hands one need at a time. Anything ambiguous becomes a question anyone in the family can answer, and a catalog that remembers each answer cut questions per shop from about 40 to 4. It runs on Claude Opus because a wrong substitution costs more than tokens.

    • AI agent design
    • Human-in-the-loop design
    • Cost & latency tradeoffs
    • Systems thinking
  • YarJobs coaches a job seeker to the job they actually want, then scans and scores postings against it.

    Helps people who don't know what's next in their careers. A coaching agent holds a real conversation that draws out deal-breakers, pay floor and commute limits, and turns them into a structured profile; I iterated the prompt with five testers. A background worker then searches job boards through MCP tools, and a scoring stage rates what survives and tracks each posting from spotted to hired. Both people who ran a full search with it were hired.

    • Discovery
    • LLM evaluation
    • MCP servers & tools
    • AI agent design
  • YarMail turns incoming email into finished work: calendar events, tasks, filed documents, and tracked orders.

    Turns email across three mailboxes into finished work, not just a cleaner inbox. The model proposes filtering rules in bulk for me to approve, and routed mail sets off actions in the other apps: calendar events, extracted tasks, filed documents, tracked orders and logged receipts. Order confirmations go to YarMeals, and mailed-in documents go to YarDocs.

    • Systems thinking
    • Human-in-the-loop design
    • AI agent design
  • YarDocs reads, names, and files every document the household receives, with nothing waiting on a person.

    Ended the folder of 800 files all named like 74249826.pdf. One inbox takes scans and forwards. Each document is read, summarized, given a consistent name and filed to the right OneDrive folder, with nothing waiting on me to decide where it goes. A $150 token bill early on taught me to choose the model per step by what it costs.

    • AI agent design
    • Cost & latency tradeoffs
    • Vercel & Azure

Experience

PayPal — Senior Technical Product Manager

Austin, TX · 2022–present

  • LLM evaluation
  • Experimentation & A/B testing
  • Systems thinking
  • Consensus building
  • Product strategy

Service Experience Platform Product: tooling and platforms for a global 9,000-person contact center serving 400M+ customers across regions and regulatory regimes. Telephony, workforce management, knowledge management, case work across every channel, and the AI agent layer on top of customer support.

  • Spearheaded real-time AI agent layer on live contacts. The summarization agent is live and cuts handle time on transferred contacts by about 20 seconds each; coaching and risk-intervention agents are in pilot. I write the prompts, define each agent's KPIs, and run the A/B and outcome evals against contact history.
  • Cut escalation-path changes across 100+ work queues from over a sprint of code releases to same-day configuration that business users own.
  • Raised merchant outreach campaign productivity over 9% by moving outbound call campaigns from shared spreadsheets into routable, auditable workflows with real analytics.
  • Contributed to the workflow primitives that break a customer intent into focused micro-workflows a service agent runs one at a time to a defined resolution.
  • Shipped the tooling service agents use inside those workflows: a knowledge-management widget, a customer transaction log, a coupons, vouchers, and cash-movement platform, and alerts bound to the workflow step where they apply in place of login-time "urgent" broadcasts.

Trusted Health — Senior Product Manager

Austin, TX (remote) · 2021–2022

  • Experimentation & A/B testing
  • Requirements
  • Mentoring

Hospital contingent-staffing platform at a Series B startup (Series C during my tenure), built for the pandemic nurse shortage. PM for two scrum teams: a nurse mobile app, a web front end for hospital administrators and staffing agencies, and production integrations with scheduling, timekeeping, and HRIS systems.

  • Expanded a 60-nurse pilot to a 10,000-nurse deployment across seven states, anchored by a major hospital system and innovation partner.
  • Delivered a 12% lift in fill rate and 14% lower incentive spend with the first feature set, guided by the division's first machine-learning segmentation of spend versus fill rate, which I ran.
  • Mentored two product managers in the division.

Family caregiving, practice operations, and independent product consulting

Austin / Dallas, TX · 2017–2021

  • Product strategy
  • Systems thinking

Primary caregiver for a parent through these years. Ran the operations of their family-law practice and consulted on product alongside it.

  • Led a two-sided shipping marketplace's product team from one-off consumer shipments toward repeat business accounts on the same platform.
  • Redesigned the practice's time capture, invoicing, and trust accounting, ending its recurring cash shortfalls.

athenahealth — Manager, Product Innovation

Austin, TX · 2015–2017

  • Systems thinking
  • Product design & UX
  • Discovery

Patient scheduling overhaul in a cloud EMR and revenue-cycle system used by practices from solo offices to national health systems.

  • Let patients self-book from anywhere by replacing per-office scheduling setups with one portable rule set.
  • Collapsed a typical practice's 100+ appointment types to 12, and cut new-practice scheduling implementation from three resource-intensive calls to one.
  • Raised the key provider-satisfaction metric 10% in the first poll after launch.
  • Started the move from use-case-specific code to modular APIs, microservices, and widgets, so a new practice assembled from parts rather than custom work.

Vision Web — Product Management Lead, then Application Development Manager

Austin, TX · 2013–2015

  • Product strategy
  • Requirements

EHR and practice-management product for vision care, taken over when it was struggling to reach market.

  • Cut maintenance cycles from eight weeks to one by automating a deployment that had consumed a full-time person and adding continuous integration; tripled the installed base in three months.
  • Moved the company from engineering-driven to product-led, with roadmaps, vision documents, and requirements standards; designed APIs and integration for a major managed-care provider modernizing its enterprise suite.

StudentScout (SeatonCorp) — R&D Manager, Product

Chicago, IL / Los Angeles, CA (remote) · 2010–2013

  • Systems thinking
  • Experimentation & A/B testing
  • Designed SmartCall, the outbound call-scheduling engine, and wrote its production dialer list-management code: time to first attempt from over five minutes to 28 seconds, contact rate from about 60% to over 80%, no added staff.
  • Built the analytics and data-warehouse platform the business and its clients read their numbers through, which sales used to close contracts.

Earlier career — Staff Management and PeopleScout (SeatonCorp)

  • Requirements
  • Systems thinking

Built PeopleScout's applicant tracking system from a spec I wrote as a user of the system it replaced, and Staff Management's HRIS platform, which turned standalone field offices into one network with routine back-office work moved to a central National Crew Desk.

Skills

Product: Product strategy and roadmaps · Discovery and stakeholder elicitation · Systems thinking and workflow design · Product design and UX partnership · Requirements (PRDs, epics, stories) · Prioritization and tradeoffs · Experimentation, KPIs, and A/B testing · LLM evaluation and prompt iteration · Guardrails and human-in-the-loop design · Cost and latency tradeoffs · Consensus building across business, engineering, and regulatory stakeholders · Mentoring

Technical: Claude Code and headless Claude agents · MCP servers and tools · AI agent design and isolation · Next.js, React, and React Native · TypeScript · Supabase and SQL · Vercel and Azure · GitHub Actions CI with staging promotion

Education

Texas A&M University — Computer Science Engineering