Applied AI Engineer

I build AI systems that behave the same way every time.

Then I find the ways they break and add the checks that stop bad output from quietly shipping. I've built production AI since early 2023, on top of 25 years in mortgage lending and sales operations, so I know the business problem before I write the code.

See the work Explore Trigger Ladder
Las Vegas, remote-first Python, VB.NET, Claude API Hands-on with LLMs since Nov 2022
Selected work

Systems I designed, built and run

Each one started as a real business problem. The numbers are measured, and the ones for systems still growing are floors.

Flagship

Trigger Ladder

At what rate does a book of loans become worth refinancing? Trigger Ladder steps the rate down one rung at a time across a simulated 5,000-consumer portfolio and runs every loan through agency and break-even tests, including VA IRRRL and FHA Streamline rules.

  • 17 rungs, stepping 2 points down from today's VA rate, built on public rate data
  • At the Jul 29, 2026 snapshot the book flipped at 5.625%, when 2,537 of 5,000 loans cleared every test
  • 156 automated tests, including 49 cases for FHA upfront mortgage insurance math (walk through it)
  • An AI-written morning brief that explains what moved and why
Read the project Open the live desk UFMIP calculator Source

Loans clearing at each rate. The flip rung is where more than half the book becomes refinance-ready.

Mortgage math

FHA UFMIP calculator

A public, step-by-step walkthrough of FHA's upfront mortgage insurance premium, checked rule by rule against HUD Handbook 4000.1.

Problem
On an FHA refinance, the upfront premium depends on a rate table, a 36-month refund chart for the old loan's premium, and a rounding rule when it is financed. The math is hard to check by hand.
System
A Python rules engine inside Trigger Ladder, plus a browser version that walks a sample loan in 7 steps and lets visitors run their own. The two matched field for field on 60,000 generated loans.
49 test cases 60,000 loans, 0 mismatches 8 steps, each with its rule
Try the calculator
AI pipeline

PipelineIQ

A multi-model AI pipeline with verification gates and cost tracking.

Problem
Thousands of records needed scoring and review by AI without costs climbing or quality drifting.
System
A VB.NET WinForms desktop app with a Python backend. It pulls from 3 external APIs, scores with the Claude API through the Batches API, and syncs with ClickUp. A three-tier AI review panel checks the output, with cost tracked per reviewer.
~90% cached input cost cut ~73% review cost cut 9,000+ AI-scored records 67,000+ lines 1,250+ commits
Operations platform

StreamlineIQ

An AI operations platform built for a regulated mortgage environment.

Problem
Every client contact took more than eight minutes of manual lookups, notes and follow-up.
System
Excel VBA, Python and the Claude API, connected to 6 external APIs.
8+ min → <90 s per contact 6,000+ records 65,000+ lines
Client build

DC Mad Hatter

A new website for a Chicago custom embroidery manufacturer. Solo and fixed-fee, from discovery to post-launch support.

Problem
The old WordPress install was compromised, and leads had nowhere reliable to go.
System
A clean migration to 5 production pages and 207 assets on HTTPS, with a PHP lead-capture pipeline. A silent lead-capture failure was caught and fixed the day after launch.
Live Aug 31, 2026 5 pages 207 assets
Visit dcmadhatter.com
Automation, then and now

ContCur

A VB.NET automation suite from 2010, rebuilt in 2026 as a call-audio transcription pipeline.

Original
Automation across 8+ desktop apps that ran in production for 6 years.
Rebuild
Python and VB.NET. Voice detection with Silero VAD, then two whisper.cpp instances running in parallel.
52% faster wall-clock 31/31 tests passing 1,140-line core
Process engineering

Mesa Grande Resort & Casino

A Microsoft Power Platform series for a fictional company, built on sample data.

Problem
A purchase-request process where people chase status by email and requests come back for missing details.
System
A Power BI report of the current state, then a Visio map of current and future process with a written spec. The idea: the system routes the work, so the person submits once. Next up is a SharePoint list with a Power Automate approval flow.
430 sample requests 4.7 days submit to PO ~539 hrs/yr to win back
Agent operations

A team of AI assistants

Four AI assistants working unattended, coordinated through a file-based message bus.

Problem
Hand-offs between assistants were slow and sometimes dropped without anyone noticing.
System
Scheduled headless runs with closed-loop delivery: every hand-off is confirmed, not assumed. Up to 12 models are compared side by side on a hard decision.
12 s → 1 s handoff 6.8 s → 2.6 s workflow step 12 models compared
Experience

Sales, operations and lending, now in applied AI

Most of what I build comes from problems I lived with in these jobs first.

  1. Applied AI Engineer

    Jul 2026 - Present
    Valentine Group
    • Runs a team of four AI assistants unattended, and root-caused an outage that cut hand-off time from 12 s to 1 s.
    • Rolled out a paid multi-vendor AI tool in stages, in production within two days.
    • Built and launched dcmadhatter.com for a client, end to end.

  2. Senior Sales & Client Relations Advisor

    Feb 2024 - Jul 2026
    Freedom Mortgage

    $188.8M closed across 543 loans, with a 65% gate-to-close rate.

  3. Senior Sales Operations Analyst

    2018 - 2023
    Carrington Mortgage Services

    Sales operations, analytics and operational reporting.

  4. Commercial Specialist

    2012 - 2018
    Lodge Financial

    SBA and commercial financing, with $125M in closings over 6 years.

  5. Founder

    2008 - 2012
    PageOne Media Corp

    Created the Rapid Ranking Blueprint SEO course: 8,000+ copies sold, 100% positive feedback.

  6. Mortgage Loan Originator & Net Branch Manager

    2001 - 2011
    Premier Mortgage Funding

    Built and ran a $100M production branch with 19 staff.

  7. National Corporate Sales Trainer

    1999 - 2002
    NCER

    Trained 3,300 representatives a month.

About

Domain first, then the code

I spent 25 years in mortgage lending, and before that I trained sales teams. I've closed loans, run a branch, built sales reports and sat on the phones. That's why the AI systems I build start with the work itself, not the model.

I'm self-taught. I've used LLMs hands-on since November 2022 and put them into production in early 2023. Since then I've built systems that run every day, with tests, cost tracking and the checks that tell me when something drifts.

Reliable over cleverSame input, same answer. Every rule gets a test, and every failure gets a root cause.
Cost is a featureCaching, batching and per-step cost tracking are built in from the start, not added later.
Built for the people using itSales, operations and client work taught me where the friction really is.
Contact

Let's talk about what you're building.

I'm open to applied AI, AI operations and solutions roles, especially where deep business knowledge matters as much as the model.

Email brianvalentine1@outlook.com LinkedIn in/brianvalentine9 GitHub Trigger Ladder source