Los Angeles, CA · Open to work

Joseph Heupler

AI Lead

GraphRAG agents, LangGraph automation, and LLM evals for support ops — Brio bot, Briopedia, and the data platform underneath. UC Berkeley Data & Cognitive Science.

Contact GitHub LinkedIn

Brio Water Technology

CSUN

Brio Water

NavAR

FlyOneO

Python

JavaScript

TypeScript

Airflow

pandas

NumPy

scikit-learn

Node.js

React

Flask

Firebase

Git

GitHub

GitHub Actions

Jenkins

Shopify

cat about.md

AI systems people actually use at work.

I'm the AI Lead at Brio Water, building GraphRAG agents, LangGraph automation, and the evals that prove they work — plus the data infrastructure underneath them.

Agentic AI & RAG

Brio bot and Briopedia — a GraphRAG support agent (BM25 + FAISS + knowledge graph), a LangGraph case-triage agent, and a 100-case Claude-judged eval set, with citation grounding, read-only tools, and human review

Data Pipelines & Modeling

dbt, SQL/Python ETL/ELT, and governed KPI models across CRM, ERP, and eCommerce

BI & Revenue Operations

Power BI (DAX, Power Query), Salesforce, and automated reporting for executive decision support

Systems Integration

Salesforce, Shopify, and Aircall unified into clean operational datasets

Looking for AI engineering, software engineering, and data platform roles where I can keep building agents and RAG systems that people actually use at work.

0K+

wholesale invoices with data-quality checks

less manual reporting time

cut in performance variance

failure-prediction accuracy

git log --oneline

Selected work history

Brio Water Technology

AI Lead

Support agents, Briopedia, revenue ops

Brio Water Technology

Software & Data Engineer

Pipelines, Power BI, Salesforce, dbt

UC Berkeley

Data Science & Cognitive Science

Where the modelling habits came from

AI Lead

Brio Water Technology

Sep 2025 — Present/City of Industry, CA

Brio Water

Built and run Brio bot, the support team's AI agent — FastAPI, Next.js, ChromaDB, GraphRAG, and LLM tool-calling that looks up articles and drafts customer emails, so reps answer on the spot instead of escalating
Built GraphRAG retrieval that fuses BM25 keyword search, FAISS vectors, and a Cognee knowledge graph on Kuzu, so the agent pulls the passage for the right model, part, and error code
Built a LangGraph agent that triages Salesforce cases — retrieves docs, reads product labels off customer photos with a vision model, and proposes updates behind deterministic Python gates, traced in Langfuse
Every answer quotes the document it came from; the model's tools are read-only, and a person reviews anything before it reaches a customer
Built a 100-case golden eval set with Claude as judge, scoring groundedness, correctness, and p50/p95 latency across local open-weight models — the leading 30B model stayed 96% grounded
Owns Briopedia, the git-backed knowledge base Brio bot reads from — troubleshooting flows, error codes, part compatibility, and safety certifications, with automated checks that fail the build on a bad import
Built SARIMAX and XGBoost revenue forecasting and churn scoring on a Salesforce-synced warehouse, retiring an older model that ran 17% high
Built one governed dbt data layer over Salesforce, Shopify, and Aircall with quality checks across 100K+ invoices — 70% less reporting and reconciliation time
Built QA analytics and Power BI dashboards that cut unnecessary product replacements 18–32% and product performance variance 51%

Software Engineer & Systems Analyst

NavAR

Jun 2024 — Sep 2025/Remote, USA

NavAR

Owned navigation analytics by instrumenting ARKit and beacon map lifecycle events into measurable KPIs
Built SQL ETL/ELT pipelines from Firebase into a warehouse for release-, venue-, and device-level analysis
Developed dbt models across staging, fact, and dimension layers for relocalization, map quality, and navigation reliability
Orchestrated Airflow workflows for ingestion, dbt transforms, validation, and reporting refreshes
Built dashboards and performance tooling to monitor regression risk by app version, location, and device
Integrated Jenkins and GitHub Actions with release validation and post-launch monitoring

Software Engineer Intern

FlyOneO

May 2022 — Sep 2022/Berkeley, CA

FlyOneO

Built full-stack web app with Node.js, React, and REST APIs for dynamic data-driven UIs
Developed Flask microservice for NLP sentiment analysis with NLTK
Proposed AI-driven automation workflows for local clinics — 10+ hours/week less manual communication
Integrated backend services and frontend components for end-to-end data processing

dbt build

Analytics engineering lab

Small animated sketches of the work behind the numbers: lineage and tests, governed metrics, anomaly checks, and the retrieval path under Brio bot.

cat brio-bot.md

Brio bot and Briopedia

QuestionA support email lands, often a photo or a PDF.
RetrieveHybrid search over the knowledge base — keyword and vector.
GroundCandidate passages are cited before anything is written.
AnswerThe model proposes; deterministic code decides what is saved.

Brio bot

The internal assistant Brio's support team uses to answer product questions in plain English from Brio's own documentation. FastAPI backend, Next.js UI, and GraphRAG retrieval that fuses BM25, FAISS vectors, and a Cognee knowledge graph on Kuzu — plus a LangGraph agent that triages Salesforce cases and reads product labels off customer photos with a vision model.

Guardrails

Every answer quotes the document it came from. The model's tools are read-only, deterministic Python gates decide what may be written to Salesforce, and a person reviews anything before it reaches a customer. Models are scored on a 100-case golden eval set with Claude as judge — the leading local 30B model stayed 96% grounded — and every run is traced in Langfuse.

Briopedia

The git-backed knowledge base Brio bot reads from: troubleshooting flows by model, error codes, part compatibility, and safety certifications, with automated checks that fail the build on a bad import. Multi-user accounts, local (Ollama) or cloud (Claude) LLM, and per-person chat history.

01Data Modeling
02Pipelines & Orchestration
03BI & Reporting
04AI & LLM
05Business Systems

Data Modeling
dbt · Star Schema · Fact & Dimension Models · Staging Layers · Incremental Models · Slowly Changing Dimensions · Semantic Layer · KPI Governance · Data Lineage · Data Quality Tests
Pipelines & Orchestration
ETL/ELT · Apache Airflow · Batch Ingestion · API Ingestion · Scheduled Jobs · Data Validation · Monitoring & Alerting · Reconciliation · Power Automate
BI & Reporting
Power BI · DAX · Power Query · Dashboard Design · Executive Reporting · Backlog Reporting · Revenue Reporting · QA Analytics · Excel Online · Ad-hoc Analysis
AI & LLM
GraphRAG · Knowledge Graphs · Hybrid Retrieval · LangGraph · LLM Evals · Langfuse · Vision Models · LLM APIs · RAG · Embeddings · Vector Search · Prompt Engineering · Knowledge Bases · AI-assisted Workflows
Business Systems
Salesforce · Shopify · Aircall · CRM Operations · Order-to-Cash · Revenue Operations · Systems Integration

Brio botBriopediaRAGdbtPower BIRevenue OperationsData Quality TestsApache AirflowSalesforceVector SearchKPI GovernanceCI/CD

ls ~/projects

Selected work

Automated Wholesale Report Splitter

Splits a single consolidated wholesale Excel report into one workbook per sales rep, so distribution stops being a manual copy-paste job every cycle.

PythonopenpyxlExcelOps Automation

Briopedia RAG Engine

RAG-grounded support agent backend over git-backed Brio product manuals, wiring FastAPI, Chroma vector retrieval, and automated citation checking.

PythonFastAPIChromaDBRAGLLM

Brio RMA Defect Classifier

NLP classification pipeline sorting dispenser warranty returns and support tickets against ERP serial numbers, so recurring manufacturing defects surface by product line.

PythonFastAPINLPscikit-learn

Predictive Filter Replacements

Time-series ML pipeline predicting water filtration cartridge exhaustion dates from Shopify D2C purchase patterns and dispenser usage telemetry.

PythonXGBoostTime SeriesShopify API

Sales ML Forecasting Demo

Implements and compares four forecasting approaches on synthetic data that preserves real seasonality and trend — the business figures stay confidential, the technique doesn't.

PythonProphetARIMAForecasting

Proactive Maintenance Predictor

Technical stack

Languages & Query

Data Modeling

Pipelines & Orchestration

BI & Reporting

Business Systems

Analytics & ML

AI & LLM

Engineering & Platform

cat education.md

Academic background

University of California, Berkeley

B.A., Data Science and Cognitive Science · Aug 2019 — Aug 2023

Pasadena City College

Associate's, Natural Sciences & Engineering · 2016 — 2019 (Prior studies before transferring to UC Berkeley)

California State University, Northridge

Probability and Statistics coursework · Jan 2024 — May 2024 (Post-graduate continuing coursework (single course))

./say-hello.sh

Let's build something precise.

Open to software engineering, data engineering, and business systems roles.

joseph.heupler@gmail.com jheupler@berkeley.edu (510) 213-8712

GitHub LinkedIn Site