Open to full-time roles · Dec 2026 · San Francisco

Elliot Overhiser

Cal Poly SLO industrial engineer. I build AI workflows and test them before anyone relies on them.

Looking forAI deploymentSolutions / implementationOperations / BizOps
Email meResume (PDF)LinkedIn
01 · inputMessy inputcustomer PO, call, request02 · agentAgent checksitems…checking✓ okpricing…checking✓ okstock…checking✓ okcredit…checking⚠ flag03 · outputVerified outputreport with evidence, flags, next action3 ✓1 ⚠
200+

test orders run against an AI agent I built inside an ERP

5-agent

pipeline I designed that writes requirements docs: 3 drafters, a blind judge, a verifier

~$840K/yr

projected savings from an AI voice agent I modeled for a health network

35.9%

shorter patient wait times in my call-center simulation

300-person

show I put on as a Juke co-founder, with 250+ presale tickets

01 — Background

Operator and builder history

  1. Summer 2026

    Internship at RSM

    AI agentsERP
    Business Applications Intern, ERP
    • Built a PO verification agent
    • Built a multi-agent BRD generator
    → Cases 01–02
  2. 2025–26

    Open Machine

    AI productsQA
    Agentic AI product testing
    • Tested and reviewed agentic AI products before and after launch
    • Feedback went to product teams at Google Labs, Anthropic, Perplexity and GenSpark
  3. 2025–

    Offrd

    Voice AIFounder
    Co-founder
    • AI-driven internship matching with outbound interview calls
    → Case 04
  4. 2024–25

    Juke

    EventsBDFounder
    Co-founder
    • App connecting venues, artists and fans
    • Ran media, business development and events
    • Put on a 300-person show with 250+ presale tickets
  5. 2021–24

    The Pennington Crew

    OperationsFinance
    Ran a 10-person landscaping business for 3 years
    • Did every estimate, plus invoicing and the books
    • Tracked crew hours and job margins
  6. 2023

    sivvie

    Founder
    Co-founder
    • AI market-research startup

Education

Cal Poly

Industrial Engineering

Dec 2026

Toolkit

Claude Coden8nAI agent designERP systemsAnyLogicExcel financial modelingElevenLabsTwilioMySQL

02 — Work

Selected work

Case 01InternshipSummer 2026ERP / AI agent

PO Verification Agent: AI order checks inside an ERP

Flow · illustrativeCase 01
Customer POlines inItem checkitem exists✓ okPrice checktiered · discountdate-valid✓ okStock checkon hand✓ okCredit laddercredit check⚠ flagReportevidence per lineCustomer POlines inItem checkitem exists✓ okPrice checktiered · discount · date-valid✓ okStock checkon hand✓ okCredit laddercredit check⚠ flagReportevidence per line
Problem

The ERP's built-in sales order agent builds orders at system prices but doesn't flag when a customer's PO disagrees with them. Wrong prices, missing discounts and over-limit customers slip through.

What I built

A read-only AI agent that reads a customer PO and verifies every line against the ERP.

  • Checks that items exist, applies tiered, discount and date-valid pricing, checks stock on hand, and runs a credit check.
  • Every check cites the evidence it found.
  • Took it through more than 20 instruction versions and built regression suites with answer keys.
  • Documented 12 capabilities and 9 hard limits of the platform.
Result
200+
test POs
87%
first-pass accuracy
0
silent wrong answers
0
writes to the ledger

Built during an internship. Demo is illustrative; no company or client data shown.

Case 02InternshipSummer 2026Multi-agent system

Master BRD Generator: discovery notes in, developer-ready requirements doc out

Pipeline · generate modeCase 02
notestranscriptscreenshotsdraftsDrafter Aclient angleDrafter Bdeveloper angleDrafter Ctester angleBlind judgescores A/B/CBuildWord templateVerifiernever saw draftsFinal BRD.docxfix & re-verifyapproved lessonslearnings.mdnext runnotestranscriptscreenshotsdraftsDrafter AclientDrafter BdeveloperDrafter CtesterBlind judgescores A/B/CBuildWord templateVerifiernever saw draftsFinal BRD.docxfix & re-verifyapproved lessonslearnings.mdnext run
Problem

Consultants turn messy discovery material (notes, meeting transcripts, screenshots, drafts from other tools) into Business Requirement Documents by hand. It's slow, details get retyped or lost, and unknowns block the document.

BRD · example (fictional)Prepared by: consultant

Sales Order Aging Column

Client: Harbor Tile Co. (fictional)
FR-01Add an Age (Days) column to the Sales Order list.
FR-02Age = today minus order date.
FR-03Highlight orders older than 30 days.
FR-04Include orders in any status.
[TO CONFIRM: include closed orders?]
What I built

A Claude skill with two modes. Generate writes a new BRD. Amend makes a scoped change to an approved BRD without touching the rest.

  • Generate runs an 8-step pipeline with 5 agents: three drafters write every section in parallel from different angles (client, developer, tester), a blind judge scores the drafts labeled only A/B/C and picks the best content section by section, and an independent verifier that never saw the drafts renders the finished Word file and checks it against a source-tagged requirements inventory.
  • It asks every missing question in one round and never asks something the sources already answer. Genuine unknowns get drafted anyway and flagged as [TO CONFIRM].
  • It edits the company's Word template directly, so branding and structure stay intact.
  • It has 116 ERP table and page IDs verified against the vendor's documentation.
  • It keeps a learning loop: after each run it proposes lessons, and only the ones the user approves carry forward. No client facts are ever stored.
How it works (5 details)
  • Generate runs an 8-step pipeline with 5 agents: three drafters write every section in parallel from different angles (client, developer, tester), a blind judge scores the drafts labeled only A/B/C and picks the best content section by section, and an independent verifier that never saw the drafts renders the finished Word file and checks it against a source-tagged requirements inventory.
  • It asks every missing question in one round and never asks something the sources already answer. Genuine unknowns get drafted anyway and flagged as [TO CONFIRM].
  • It edits the company's Word template directly, so branding and structure stay intact.
  • It has 116 ERP table and page IDs verified against the vendor's documentation.
  • It keeps a learning loop: after each run it proposes lessons, and only the ones the user approves carry forward. No client facts are ever stored.
Result
5
agents
3
competing drafts per section
116
verified ERP objects
0
re-asked questions

Built during an internship. Example content is fictional; no company template or client data shown.

Case 03Senior project2025–26Operations research

AI voice agent evaluation: Community Health Centers of the Central Coast

Problem

A 28-clinic health network's patient call center had long holds and dropped calls on appointment scheduling. Leadership wanted to know whether an AI voice agent would actually help.

What I built

A model of the call center, current state vs. AI-assisted state.

  • Cost model in Excel.
  • AnyLogic discrete-event simulation built from real call-center data.
  • Presented the findings to CHC leadership.
Simulation · before / afterCase 03
Wait timecurrent100AI-assisted−35.9%Abandoned callscurrent100AI-assisted−16.7%Indexed: current state = 100
Result
~$840K/yr
projected savings
35.9%
lower wait time
16.7%
fewer abandoned calls
42.4%
smaller queue
Case 04Founder2025–presentAI product

Offrd: outbound AI interview agent

Problem

Students applying for internships rarely get a real first conversation. Recruiters can't screen everyone.

What I built

An outbound voice interview agent using ElevenLabs Conversational AI and Twilio.

  • Each call is personalized to the candidate.
  • Structured data is extracted after the call, so results are comparable across candidates.
Call flow · sample dataCase 04
TwiliodialElevenLabsagentTranscriptfull callJSONstructuredTwiliodialElevenLabsagentTranscriptfull callJSONstructured
{
  "candidate": "Sample Candidate",
  "role": "Ops Intern",
  "availability": "Summer",
  "interest_areas": ["supply chain", "analytics"],
  "follow_up": true
}

03 — How I operate

How I operate

  1. 01Scope
  2. 02Map the workflow
  3. 03Build and test
  4. 04Measure
01

I've done the deployment work.

Scoped, built and stress-tested a production-style AI agent inside an ERP, including the unglamorous parts: permissions, edge cases, documentation.

  • 200+ test orders with answer keys
  • 9 platform limits documented alongside 12 capabilities
02

I think in systems.

Industrial engineering training means I model the process before I touch it. Simulation, cost models, workflow maps.

  • CHC: turned "should we use AI?" into a cost model and a simulation
  • PO verification mapped into 12 steps; BRD writing into an 8-step pipeline
03

I've run things.

A 10-person business, a 300-person event, two startups. I'm comfortable owning outcomes without a playbook.

  • 3 years running The Pennington Crew
  • 300-person Juke show, 250+ presale tickets