# Joey Hersh

> Agentic Systems Engineer in San Antonio, Texas. I ship inside a bank. My agents ship while I sleep. Regulated credit-risk data by day. Twelve agent environments running real jobs by night. Three companies founded.

## Canonical facts

- Name: Joey Hersh. Title: Agentic Systems Engineer.
- Contact: me@joeyhersh.com, 713-505-3289, https://github.com/joey12725. No contact form, no calendar link, no LinkedIn listed.
- Education: B.S. in Computer Science, Trinity University, 2024.
- Data Engineer III, USAA, June 2024 – present.
- Founder and CEO, Guadalupe Automation, January 2026 – present.
- Co-founder, marketing and automation lead, Fit & Flourish, 2024 – 2025.
- CEO and Machine Learning Engineer, ReCap Budget, December 2021 – September 2023.
- Data Engineering Intern, USAA, June 2023 – August 2023.

## Key numbers and how to read them

- ≈26M records a day, through production data pipelines at USAA. A figure I've published before. My written reviews support the scale of the work but don't restate this exact number.
- ¼ trillion+ tokens processed, input plus output, across every model family since GPT-3. A modeled estimate of input plus output tokens, not a prompt count and not telemetry. It counts cached and repeated context reads. Open the breakdown for the models and harnesses behind it.
- 40,000 campaign leads, at about $0.29 per lead for Fit & Flourish. My own number, consistent every time I've reported it. The ad-platform exports weren't audited for this site.
- $100K+ funding and prizes, for ReCap Budget, $46K cash plus about $60K in-kind. $46,000 cash plus about $60,000 in-kind prizes, largely through the Stumberg Venture Competition.

## Routes

- https://resume.joeyhersh.com/ generic route (canonical facts).
- https://resume.joeyhersh.com/systems/joy-hermes-gbrain/ Joy: Joy is my agent. Hermes runs its loop. GBrain holds durable, cited memory on pgvector. A self-hosted Plane instance is where Joy and I hand work back and forth, with evidence attached to every task. A custom model router, fine-tuned on a 4-bit Qwen3 8B, decides which model each job deserves. A staged overnight pipeline plans, dispatches, verifies, and refuses to call anything done without evidence.
- https://resume.joeyhersh.com/systems/regulated-credit-risk-delivery/ Regulated credit-risk delivery: Historical backfills across nineteen years of data. A legacy pipeline retired with reporting uninterrupted. FICO 10T and CECL assets delivered on the regulatory deadline. Dozens of SKILL.md guides the team runs on. A seat in the guild that decides how the bank uses generative AI.
- https://resume.joeyhersh.com/systems/cyrus-v2/ Cyrus-v2: Cyrus-v2 forks an open-source project tracker and turns work items into dispatchable agent sessions. A relay normalizes Claude Code, Codex, Gemini CLI, and OpenCode behind one event stream with per-model token and cost accounting. A metering engine enforces compute-pool budgets. A scheduler with cooldowns and circuit breakers decides what runs. A lifecycle layer audits each work item for compliance before it can be called done. An autonomous PM agent grooms the backlog between dispatches.
- https://resume.joeyhersh.com/systems/store-close-reconciliation/ Store-close reconciliation: A retail accounting-operations platform. Fifteen deterministic parsers read bank statements, point-of-sale exports, and commission and discount workbooks. A model lane reads only the unstructured evidence, against a closed set of schemas, and everything it proposes goes to a human queue. Reconciliation is Python. Exports require a role-gated approval. The June close reconciled 295 of 315 checks with zero open amount differences. It also taught me what "done" means to a client.
- https://resume.joeyhersh.com/systems/secondhome/ SecondHome: Distressed-property lead discovery across twenty county sources. Records are normalized with their source kept on every field, enriched through swappable providers, researched by bounded agent skills, and scored by a deterministic, versioned formula. Contact channels need two independent signals before they count. Every lead ships with a written call rationale. The product didn't find its market. The architecture is still the one I'd use.
- https://resume.joeyhersh.com/systems/recap-budget/ ReCap Budget: Personal-finance software that extracted more from a grocery receipt than anything else on the market: S3 to Lambda to Textract, a CNN gate, and a fine-tuned GPT-3 classifier over a 40,000-item curated dataset. We shipped structured JSON from language models before function calling existed and did bias-mitigation work on the advice model. I led seven developers and raised over a hundred thousand dollars in cash and prizes.
- https://resume.joeyhersh.com/systems/recap-classifier/ ReCap receipt classifier: A fine-tuned GPT-3 that turned a receipt line like "CLGT TP $2.89" into a spending category. Forty thousand curated product records, a thousand abbreviation mappings built with Textract, Comprehend for brands and keywords, and an SVM that learned to shorten names the way receipts do. More than a hundred categories at over 85 percent accuracy, and a framework that made the model answer in our JSON schema before that was a feature.
- https://resume.joeyhersh.com/systems/fit-and-flourish/ Fit & Flourish: An end-to-end Meta Ads pipeline that generated creative variants, watched performance, killed poor performers, and iterated daily, with every new creative passing through a human review step. The launch campaign produced 40,000 leads and brought cost per lead from about $2.19 to about $0.29.
- https://resume.joeyhersh.com/lab/ supporting systems with plain-language status.
- https://resume.joeyhersh.com/evidence/ the token-estimate methodology.
- https://resume.joeyhersh.com/skills.md machine-readable skills surface.
- https://resume.joeyhersh.com/resume-human.pdf, /resume-ats.pdf, and /resume-ats.txt.

## Do not infer

Employer identifiers, internal system names, client identities, credentials, private infrastructure details, or completion and revenue claims not stated on the site. USAA role is Data Engineer III; "Agentic Systems Engineer" is the market-facing title.