VisionAirJoshua Ermert
Full-Stack Engineer
AI Software Engineering · Consulting · Analytics · Workflow Optimization
I build full-stack web apps and AI systems end to end, from data model to interface, with testing and CI baked in. Recent SDSU grad; everything below you can open, run, or read.
About
I work both sides of the decision: the analytics that turn fragmented data into answers, and the secure engineering that puts those answers into production.
- Focus
- Full-stack development · LLM integration · data & analytics · CI/CD & testing.
My path wasn't a straight line, and that's where the value is. It started in defensive security (an NSA/NSF-funded cybersecurity camp, then a City of San Diego internship), ran through a Management Information Systems degree with a Computer Science minor at San Diego State, and picked up a designer's eye during a 250-hour internship in Rome. MIS taught me how systems serve a business; CS gave me the depth to build them.
Since 2024 I have run an independent AI consulting practice. The anchor is a production fitness-coaching platform I own end to end; around it sit a source-verified research library for a physical-therapy professional and a governed operations system for a small business. At Celltrion, inside a regulated pharmaceutical environment, I built the analytics layer a commercial-operations team relied on every week: a Power BI master-data report used nationwide, with anomaly detection over more than ten thousand records.
The work that follows keeps a consistent shape: production applications with testing and CI, LLM features that stay grounded in real sources, and data work that knows when a headline number is misleading. The claims stay checkable: live links and public repos where they exist, honest labels where the work is private.
Work
Secure web apps by construction
Security is a discipline I build in from the start, not a patch applied later. Strict Content-Security-Policy and HTTP security headers, zero secrets in the client bundle, least-privilege by default, and untrusted input (model output included) handled behind explicit boundaries. The aim is to remove whole classes of vulnerability by design rather than chase them after the fact.
Responsible LLM systems
I wire LLMs into real workflows the way you would run a system you are accountable for. Outputs are grounded in verified sources instead of left to fabricate, model output is treated as untrusted behind an explicit tool allowlist, and a human stays in the loop for anything irreversible. Behavior holds up under real, messy use.
Full-stack delivery, end to end
Schema to interface to deploy: typed React and Next.js products backed by real data and state, shipped through automated pipelines (GitHub Actions running lint, tests, build, and secret scanning) with pinned, lockfile-enforced dependencies. The work below runs from a live decision-support app to governed-AI systems delivered for real clients, and every one of them is a full build.
VisionAir
METMET
Built · deploy pendingA nutrition and training platform built end to end for a working strength coach's practice.
- TypeScript
- Next.js 15
- React 19
- Prisma + Neon Postgres
- Clerk (invite-only auth)
- Vitest · 319 tests
Details
Mission ControlMission Control
Private tooling · in active useOrchestration that runs AI coding agents in parallel behind automated quality gates.
- Next.js
- React
- TypeScript
- Git worktrees
- Automated test + build gates
Details
HydrOSHydrOS
Research simulation · localA reinforcement-learning simulation of a multi-stage water-filtration system, validated before hardware.
- Python
- PPO / Stable Baselines3
- Gymnasium
- pytest
Details
CDC Diabetes-Risk Analytics POCCDC Diabetes-Risk Analytics POC
Academic POCAn academic ML proof-of-concept classifying three-class diabetes risk on a large CDC public-health dataset.
- R (nnet, class, rpart)
- Python (pandas, matplotlib, python-pptx)
Details
Validated Research LibraryValidated Research Library
Independent consulting · deliveredA governed-AI knowledge system turning unstructured expertise into a source-verified, queryable library.
- LLM orchestration
- Source-grounding & traceability
- Per-record metadata schema
- Human-in-the-loop review
- Validation suite
Details
Governed Operations SystemGoverned Operations System
Independent consulting · deliveredA queryable operations knowledge library for a small-business client, with a human committing every change.
- LLM orchestration
- Conversation-corpus mining
- Request routing
- Controlled structure
- Git-based tuning
Details
AlignFlowAlignFlow
Design-stage prototypeA multi-source data-quality / master-data-management engine for reconciling records into one trusted view.
- Python
- pandas
- fuzzy matching
- Excel ingestion
- Power BI
Details
Experience
Independent AI Consultant
7/2024 – Present · San Diego, CASince 2024 I have run an independent practice building software for clients who needed a working system, not a demo. The anchor is a production fitness-coaching application I own end to end: a Next.js and TypeScript monorepo over a 49-model Postgres schema, with decision engines for progression, scheduling, and nutrition, held to more than 200 automated tests and a human coach reviewing every consequential call. Around it I built a source-grounded content system for a physical-therapy practice that ran 30 days in production with zero fabricated references, and a conversational intake system for a private chef that turned cold inquiries into paying clients and cut a core workflow by about a third. The common thread is trust: taking a non-technical goal and turning it into a system the client can rely on and run.
Celltrion
Summer 2025At Celltrion I built the analytics and decision-support layer for a commercial-operations team inside a regulated pharmaceutical environment. The centerpiece was a production master-data report in Power BI, split into an operations view and a sales-validation view, that teams leaned on every week across all 50 states and the U.S. territories. I also stood up anomaly detection over ten-thousand-plus records with a triage workflow that sorted issues by frequency and impact, and I scoped an Azure OpenAI integration with the secret-handling and human-review gating a regulated setting demands. It was where I learned what production really means when a real business depends on the numbers being right.
Iperdesign
Summer 2023 · Rome, ItalyA 250-hour design internship in Rome, where I designed and built three client websites in WordPress and Figma on an international team. Working across languages and stakeholders taught me the part most engineers skip: how a system communicates its purpose is part of the engineering, not a coat of paint. I still design with that in mind.
City of San Diego
2022My first technical role, a cybersecurity internship where I worked through Linux and defensive-security fundamentals and helped run the learning operations for the intern cohort. It is where a habit started that never left: think about how a system fails before you think about how it ships. That instinct now shows up in every threat model and test suite I write.
Education
B.S. Management Information Systems with a CS minor, honors, cum laude. San Diego State University, Weber Honors College. Graduated May 2026.
Work with me
I build software for people who have a business to run.
I work with businesses that have outgrown their tools: manual processes, disconnected spreadsheets, or an idea that needs to become real software. I handle the full build myself and deliver a system you can run your operation on. The work above is the record.