Principal AI Systems Engineer

Applied AI systems, shipped — from prototype to production.

I'm Jase Stewart. I take AI the whole distance — from a promising prototype to a system a business can trust with a real decision. A decade forecasting risk in regulated finance, where every model had to survive an audit, taught me the difference between a demo and something that holds up in production.

Book a working sessionSee the work
12+ yrs
Forecasting & governed analytics in finance
34+ tools
Routed by the Selin agent platform
3 days → 6 min
One workflow, automated at Wells Fargo

The work

Eight systems, not eight slides.

Each of these I designed, built, tested, and operate — solving a problem the whole way down, from the data layer to the guardrails. Hover any card for the details that matter to the people who'd rely on it.

Agentic developer platformIn production

Automated Coding Engine

Repeatable, inspectable delivery — not unattended code generation.

A self-hosted software-engineering system that coordinates implementation, independent review, critique, arbitration, security testing, regression checks, telemetry, and human escalation. I use it to build and modify real applications.

  • Implementation → review → critique → arbitration
  • Security testing, regression checks, telemetry
  • Human escalation on consequential steps
Multi-agentDASTSandboxReliability
Agentic assistant & integration platformIn production

Selin

Routes 34+ tools — with permission checks on the risky ones.

A local-first assistant and integration platform combining reusable tool interfaces, model routing, retrieval and memory, identity checks, capability-based authorization, and human approval for guarded actions.

  • 34+ tools across connected workflows
  • Model routing + retrieval + memory
  • Capability-based auth; human approval for guarded actions
AgentsTool useRoutingMemory
Identity, policy & live defenseIn production

Security Layer

Layered defense for live web properties.

A layered defense for two live web properties: request signatures, ML detection, cross-site campaign monitoring, Cloudflare edge rules, incident alerts, and reviewed-feedback retraining. I distinguish confirmed enforcement actions from observed request outcomes.

  • Request signatures + ML detection + edge rules
  • Cross-site campaign monitoring & incident alerts
  • One campaign: 20 source IPs auto-blocked; no sensitive files served
AuthML detectionCloudflareAudit
Applied ML platformIn production

SelinAI ML Workbench

Dataset ingestion to evaluation, behind an API.

ML tooling for dataset ingestion, training and inference, model evaluation, and API access — the reusable core behind the applied-ML work, exposed over REST and an MCP interface.

  • Ingestion, training, inference, evaluation
  • REST + MCP access
  • Model evaluation built in
MLMCPEvaluationAPI
Applied ML · MarketsIn active use

MarketPulse

Market intelligence — modeled, alerted, backtested. Not live brokerage.

A Python market-intelligence workflow integrating market and macro data, modeling, alerts, backtesting, and paper trading — explicitly not live brokerage execution.

  • Market + macro data, modeling, alerts
  • Walk-forward backtesting
  • Paper trading, not live execution
PythonForecastingBacktestingAlerts
Retrieval-driven applicationIn progress

Legal Analyzer

Keeps legal authority separate from client evidence.

A legal-analysis system with metadata-aware hybrid retrieval and access controls that keep legal authority separate from client evidence. GCP deployment remains in progress.

  • Authority separated from client evidence
  • Metadata-aware hybrid retrieval + access controls
  • GCP deployment in progress
RAGLegalAccess controlHybrid search
Adaptive tutoringIn active use

Praeceptor Ultimus

A local-first adaptive tutor that teaches from your own material.

A local-first adaptive tutor with document ingestion, citation-aware retrieval, local/cloud model routing, and interactive study tools — all running on your own material.

  • Document ingestion + citation-aware retrieval
  • Local / cloud model routing
  • Interactive study tools
RAGLocal-firstCitationsTutoring
Multimodal engineeringLive

SelinAI Studio

Local-first image/video and voice/music production.

Multimodal production workflows with Python APIs, MCP interfaces, authenticated services, and regression tests. Its identity-consistent image pipeline combines character-specific LoRA adapters, Blender/IK pose drivers, diffusion conditioning, compositing, computer-vision quality gates, and experiment logging.

  • Identity-consistent pipeline: LoRA + Blender/IK + diffusion
  • Computer-vision quality gates + experiment logging
  • Python APIs, MCP, authenticated services
VoiceDiffusionBlenderPipeline
SaaS · CoachingIn production

Ultimum Sales Trainer

Multi-tenant AI sales coaching, safely isolated.

A multi-tenant AI sales-coaching product with tenant-isolated knowledge bases, pgvector + Postgres full-text hybrid retrieval, PII scanning, an enforced AI egress policy, and bilingual retrieval.

  • Tenant-isolated knowledge bases
  • Hybrid retrieval (pgvector + FTS)
  • PII scanning + enforced egress
Multi-tenantHybrid searchPIIBilingual

Live demos

Try them in the browser

Real running builds, embedded live below. Open either one full-screen in its own tab.

Drunk Interpreter

Type a line, pick a persona, get it slurred back.

open full ↗

This public build uses the deterministic styling engine; the agentic Claude/Grok styling runs on the private build. Voice input pauses while the GPU is busy.

MarketPulse

Indicators, an autonomous paper portfolio, and sentiment-scored news.

open full ↗

Read-only dashboard — tracked equities, macro / commodity / crypto series, the $10k paper simulator, and scored headlines. Market intelligence, not brokerage.

Corporate Travel Copilot

Plain-English trip → a policy-compliant itinerary, savings vs baseline, and an inspectable reasoning trace.

open full ↗

Prototype with mock inventory — books nothing. Deterministic policy core (cabin/airfare/vendor/hotel rules) with an agentic parse + reasoning layer and a deterministic fallback. Built end-to-end by the SelinAI engine under a 100%-coverage / DAST / independent-review contract.

Approach

Evidence over vibes.

The same three tests on every system I build — the reason each one survives contact with real users and real audits.

Inspectable

Narrow tools, explicit permissions, capability-based authorization — decisions and access you can look at, not a black box holding the keys to everything.

Measured

Define the workflow, then measure the result against it. Evidence over vibes, and a test for the paths where it breaks — not just the one where it works.

Accountable

A person stays on the hook for consequential actions. Guarded operations wait for human approval; isolated execution contains the rest.

About

AI that does real work — and shows its work.

I build AI systems that do real work — and make their decisions, permissions, and failure modes inspectable. My background pairs more than a decade of forecasting and governed analytics in financial services with hands-on development of self-hosted agentic platforms, ML services, developer tooling, security infrastructure, and multimodal applications.

It's an unusual pairing — regulated-finance discipline and full-stack, self-hosted AI engineering — and that's the point. Everything in the work above I designed, built, and operate end to end.

If you have a problem worth solving with AI, I'd like to hear it.

At a glance

Title
Principal AI Systems Engineer
Focus
Agentic platforms, developer tooling, applied ML
Based
Des Moines, Iowa · Fully remote (U.S. authorized)
Engagements
Working sessions, builds, and advisory

Experience

Where the rigor comes from.

More than a decade of governed analytics in financial services, then full-stack AI engineering — same standards, new tools.

  1. Wells Fargo

    Senior Data Scientist / Analytics Lead

    Production forecasting and governed analytics for consumer lending. Automated one recurring workflow from roughly three days to six minutes.

    2013–2026

  2. UltimumGroup

    Co-Founder & Principal AI Architect

    Built agentic applications and reusable AI engineering infrastructure.

    2022–Jul 2026

  3. Racing Chemistry

    Statistical Research & Systems Evaluation

    Statistical research and evaluation of measurement systems.

    2006–2011

Education

  • M.S. Statistics — Texas A&M University
  • M.S. Business Analytics — Drake University
  • B.L.S. — Iowa State University

Technologies

  • Python
  • TypeScript / Node.js
  • SQL
  • Docker
  • Linux
  • MCP
  • Local & multi-provider LLMs
  • Hybrid RAG
  • Evaluation
  • scikit-learn
  • XGBoost
  • Playwright
  • gVisor
  • PostgreSQL / SQLite
  • Cloudflare
  • RS256 / JWKS
  • OIDC / SAML
  • Blender
  • ComfyUI

Let's build

Have a problem worth solving with AI?

Book a working session. Verify your email, pick a time, and we'll talk through what's real, what's hard, and what's worth building.

Book a sessionEmail me