Open to
Full-time or contract, in the US or in Seoul.
Full-time
Executive product and technology roles: CPTO, CPO, VP or Head of Product, CTO.
US remote, Austin, or relocation for the right seat. APAC roles from Seoul.
Contract and fractional
Fractional CPO or CPTO, AI product strategy, pricing and data products, for fintech and financial services.
In the US and in Seoul. U.S. citizen, no sponsorship needed; Korean F-6 residence visa with full work authorization.
Why now: Mitra is built and in the hands of its President, who is taking it to market. That was always the plan. George is looking for the next challenge.
Email GeorgeGeorge Neal spent twenty years building data-driven products for some of the largest financial institutions in America: credit products and pricing inside a bank, then product and technology leadership at companies pricing over a trillion dollars a year in loans. He learned one thing above everything else: the biggest competitive advantages go to whoever figures out the new technology first and builds around it seriously, not decoratively.
In 2024 he stepped out of corporate leadership to test that idea against AI directly. Not with slide decks, with working systems that have to earn their keep. Some are live businesses. Some are running experiments. One was shut down on purpose. All of them are below, with the record that came before them. Everything on this page carries an honest status.
The record
Anchors, not adjectives. Each one survives a reference check.
AI customers paid for
Bank of America and TD Bank called the AI assistant he built "the single reason we selected the platform," paying over $10M a year combined.
A turnaround with numbers
Recruited by private equity owners into a business with under a year of runway and 60% churn. A year later it ran at 103% net revenue retention.
Zero to one, at scale
Designed a fintech product that generated $20M in its first year; built a data products business from one person to $5M+ revenue and a team of twenty.
Credit and lending, from every seat
Credit products P&L at a regional bank, RAROC pricing bank-wide, a seat on the commercial credit committee; later, product leadership for platforms pricing over $1 trillion a year in loans.
Teams that stay
A data products team grown from one to twenty with voluntary turnover at 5% or less; at a turnaround, voluntary turnover cut from 50% to under 5%. References include people who worked for him and signed up to do it again.
In print and on stage
Author, "Model Validation," The Definitive Guide to CDOs (Risk Books). Speaker at Bank.AI and twice at CDAO Financial Services. Current research: a transformer-based market-regime model with 14.5 years of out-of-sample validation.
Career
Newest first.
- 2025 to presentCPTO & Co-Founder · Mitra AI
- 2024 to presentCo-Founder & CPO · The Coaching Record
- 2022 to 2023Chief Product & Technology Officer · Nomis Solutions
- 2019 to 2022VP, Data & Analytics Products · Q2 Software
- 2016 to 2019Head of Data Products / Chief Analytics Officer · PrecisionLender
- 2005 to 2016SVP Credit Products; SVP Technology & Operations; FVP Information Security · American Savings Bank
- EarlierHawaiian Electric Industries, Aon, Docucorp, UBS
Education: MBA, Quantitative Finance, and BA, Mathematics (Honors), Hawaii Pacific University.
Projects
Every effort here was built to answer a specific question. The status is the current answer.
Mitra Live
Thesis: an AI tutor that leads with connection can reach students that conventional tutoring loses.
An AI tutoring platform for middle and high school students with ADHD and learning differences, delivered in chat where they already are (KakaoTalk, Telegram, WhatsApp). Engagement first, content second. A separate Mitra for Academies line gives hagwons and schools a curriculum mastery model built from two independent assessment sources. Paying families in Korea.
What it is proving: a multi-agent teaching system (teacher, expert, safety) can hold a distracted teenager's attention session after session. The hard problems turned out to be memory, cost per turn, and safety, not pedagogy.
mitratutor.comSuyeo In development
Goal: when an AI agent acts for a person, the site it acts at can tell that a real person stands behind it, and what that person allowed.
The current venture. The person stays in charge: which agents may act for them, what each may do, until when, and a way to stop any of them. The site gets a clear answer instead of guessing whether to let the agent in.
The Coaching Record Live
Thesis: youth sports programs lose athletes because communication between coaches, parents and athletes breaks down.
A communications hub for youth athletics connecting coaches, athletes, parents and program coordinators. Coach-to-parent communication, athlete development tracking, program coordination. Bootstrapped, for-profit SaaS.
What it taught: the product is straightforward. The market is a relationship business, and AI-assisted outbound only works when a human owns the conversation that follows.
thecoachingrecord.comRubric Prototype
Thesis: hiring goes wrong because screening starts from the job description instead of the hiring manager's head.
An AI interviews the hiring manager, derives a two-layer rubric (standard job axes plus the team's own intangibles), locks and hashes it, generates the JD from it, then scores every resume per dimension with quoted evidence and a stated confidence. Built with the compliance layer in from day one: hash-chained audit log, pre-use candidate notices, adverse-impact analysis, designed against the strictest combination of NYC, California, Illinois, Colorado and EU AI Act rules.
What it is testing: whether "score evidence, not people" can be both more useful to a manager and more defensible to a regulator than the tools on the market today.
Harmonic Paper trading
Thesis: market regimes can be tokenized like language and predicted with a small transformer.
A market regime model that reduces derivatives across 68 instruments into 128 regime states using PCA and ICA decomposition, then a transformer that reads the sequence. A daily automated trader consumes the signal and runs two parallel paper books, one on high-conviction days only and one on medium and above, as a live A/B test of the conviction gate.
What it is testing: whether the signal survives contact with real fills before any capital is committed. Live mode is gated on the paper record, not on enthusiasm.
Content Factory Live
Thesis: a founder's marketing should cost a few hours every three weeks, not a few hours every day.
Mitra's English-language social distribution system. A short steering burst from the founder becomes weeks of revolving, multi-channel posts with no human in the upload loop. Two tracks (parents and academies), a single cockpit that ranks the few things only the founder can do, and a hard review gate before anything posts.
What it taught: the first attempt died at a manual CSV upload step. Automation that leaves one human handback in the loop is not automation. Rebuilt on a scheduling API and it has run since.
Trading Bot and Polybot In development
Thesis: a trader with a written creed and a memory of its own past trades gets better with time.
An options trader on a live brokerage account operating on fourteen written principles, with a chat expert and a daily runner that share one memory and a reflection system that learns from past trades. Polybot applies the same idea to prediction markets with three independent agents voting on entries, position size gated by Harmonic's daily regime signal.
What it is testing: whether discipline encoded in text beats discretion, and whether the projects above can feed each other rather than compete for attention.
Videolab In use
Thesis: short-form video production can be config-driven end to end.
A multi-company video studio. One company config drives branding, language and voice; a topic file becomes narrated, illustrated, rendered portrait video through text-to-speech, image generation, optional image-to-video, and programmatic rendering. Serves Mitra, The Coaching Record and other efforts as a shared production line.
What it taught: generation is cheap, taste is not. The human step that survived is editorial, not technical.
Autoworld Shut down
Thesis: AI agents can fully run low-margin, repeatable business models with no humans in operations.
A micro holding company that spawned autonomous portfolio companies (outbound marketing, newsletter audiences, small-business websites), each run by an AI CEO agent with a founder-level portfolio manager above it. The bet was that per-unit labor cost, not capability, was what kept these markets underserved.
What it taught: agents could produce the work, but the customer-acquisition side of a low-margin business still needed human judgment at every step. The unit economics never closed. Shut down deliberately; the operating discipline it produced carried into everything else here.
Chronicle RPG and Tabletop VTT Side projects
Thesis: the same agent and memory patterns that run a business can run a world.
Chronicle RPG is a multiplayer text RPG where every character is a trainable AI agent: one universal prompt, and a personal knowledge corpus that is the character. Tabletop VTT is a projector-first virtual tabletop for in-person D&D with imported dynamic lighting and fog of war.
Why they are here: they are where the memory and multi-agent ideas get tried without customers watching.
Portfolio
Evidence you can open, read, or ask about.
Case study: the Nomis turnaround
Private equity owners recruited George as Chief Product Officer of Nomis Solutions, an AI price-optimization platform for consumer credit at major banks, with under a year of runway and 60% churn. He moved the product from bespoke per-client builds to one standardized SaaS roadmap every client received, doubled feature output, and reached 80% roadmap attainment in six months. Product sat in every major client meeting. He stopped a product whose unit economics lost money on every seat. Voluntary turnover fell from 50% to under 5%; eNPS rose from -15 to +45. Within a year: 103% net revenue retention, the first client upsells in four years, ACV up 20%. The role grew to CPTO.
Case study: Andi at PrecisionLender
Andi was the first dedicated AI assistant for commercial lenders, giving bankers advice at the moment of decision. George led, architected, and built the machine learning behind it: real-time recommendations ranked per user on history and live deal context. Skills competed in an internal auction for scarce response slots on expected value and click-through. Every skill reported its effectiveness, and a skill whose outcomes kept missing its model was shut down. Results were validated by A/B testing and causal inference. Bank of America and TD Bank, paying over $10M a year combined, named Andi as the reason they selected the platform. It still runs inside most of the Q2 platform today.
- Published"Model Validation," Chapter 22, The Definitive Guide to CDOs (Risk Books, 2008).
- Working paper, in progressA transformer-based market-regime model: 85.9% accuracy over 14.5 years of chronological out-of-sample data, co-authored.
- TalksBank.AI (2018); CDAO Financial Services (2019, 2021).
- Live productsmitratutor.com · thecoachingrecord.com
A note on AI use. This page, like most of what George ships, was built with AI assistance. The facts on it are facts, drawn from a maintained store of roughly 250 verified claims about his career. The voice is his; the process is iterative; it is not fully automated because it has not yet earned the right to be.