Senior AI product,
for the work that has to be right.
10+ years in enterprise product · AI focus since 2023 · Atlanta · Columbia MS Applied Analytics
Foundry AI Ops is a senior fractional AI product practice for companies shipping AI in regulated, enterprise, and high-stakes contexts, where AI has to be right the first time.
Who we work with
A focused practice for teams shipping AI into contexts where shipping the wrong thing matters more than shipping fast.
- ·AI-native startups, Series A to C, shipping AI as a core product
- ·Enterprise B2B with AI on the roadmap
- ·Regulated financial services: banking, wealth, insurance, brokerage
- ·Climate and energy: verification, validation, audit-ready compliance
- ·Healthcare AI: production pipelines where output quality is non-negotiable
- ·Pre-seed and seed-stage founders building their first AI product
Regulated experience that translates to any AI work where the stakes are high, even outside compliance contexts.
Services
Four engagements: scoped, priced as ranges, selected to match the buyer's stage and risk profile.
01
4 to 8 weeks · $25,000 to $50,000
AI Product Strategy and Roadmap
Discovery, strategy, and roadmap design for AI capabilities in regulated or fiduciary contexts. Eval framework design. Rollout sequencing across shadow, canary, and A/B. KPI definition that ties to operational impact, not vanity metrics.
For: Teams who need a defensible AI roadmap before they spend engineering cycles or capital.
02
Most common engagement3 to 6 months · $11,000 to $22,000 / month
Embedded Fractional AI PM
Drop in as your AI product lead for a defined engagement. Specs, evals, guardrails, human-in-the-loop review, rollout discipline. Sprint-based delivery alongside your engineering team.
For: Teams shipping AI features without a dedicated senior PM to own the work.
03
3 to 6 months · $8,000 to $15,000 / month
Fractional CPO for 0-to-1 Builds
Drop in as your fractional Chief Product Officer for early-stage product builds. Product strategy, MVP scoping, customer discovery, hiring guidance, and the product-narrative work that goes into seed-round fundraising. Specialized in 0-to-1, not scale-up.
For: Pre-seed and seed-stage founders shipping their first product, especially those preparing for or just past their seed round.
04
2 to 3 weeks · $15,000 to $25,000
AI Audit and Eval Framework Review
Independent review of your AI product or pipeline. Eval rigor, hallucination mitigation, deterministic guardrails, failure-mode coverage, audit-readiness.
For: Teams approaching a launch, certification, or audit and wanting an external read before they ship.
Recent project · Independent AI/ML work · 2026
Prediction Market Anomaly Detection
The problem
Prediction markets move real money on real outcomes. When someone places a $97,000 bet minutes before a market resolves, that's either extraordinary conviction or information the market doesn't have yet. Distinguishing between the two, at scale, automatically, with measurable confidence, is an anomaly detection problem. I built a system to do exactly that on live Polymarket data.
What I built
A production-structured anomaly detection pipeline that ingests live trade data, scores each transaction across multiple signal dimensions, flags statistically unusual behavior, and surfaces results in a deployed dashboard, updated daily with one command.
The detection engine computes per-market z-scores across two independent signals: bet size relative to market baseline, and timing relative to market close. A third composite signal, confluence, fires when both signals are elevated simultaneously. Thresholds are set at 2.5σ for size, 2.0σ for timing, and 1.5σ on each dimension for confluence. Per-market normalization ensures a $500 bet in a $5,000 market scores appropriately differently than a $500 bet in a $10,000,000 market.
The data pipeline hits Polymarket's public API, stores raw trades and computed signals in SQLite, and exports a static JSON artifact that the dashboard reads. The full refresh covers fetch, detect, evaluate, and publish; it runs as a single shell command and pushes to production automatically via GitHub and Vercel.
The evaluation layer is deliberately honest about what it can and cannot measure. Without ground truth labels, true precision and recall are unavailable. What the system can measure: flag rate (1.8%), flagged volume concentration (48.2% of total dollar volume in 1.8% of trades), and flag distribution across signal types. The 44× ratio between mean flagged bet size ($3,680) and mean normal bet size ($84) provides strong circumstantial evidence the detector is finding real signal rather than noise.
The dashboard is a deployed Next.js application showing summary statistics, flag type breakdown, per-market flag rates, and a filterable table of flagged trades with full signal detail.
Key results
Trades analyzed
10,000
across 20 active markets
Flag rate
1.8%
Flagged volume share
48.2%
of total $ volume
Mean flagged bet
$3,680
Mean normal bet
$84
Largest flagged trade
$97,269
NBA Finals market
Confluence flags
28
large AND late simultaneously
What I documented honestly
A system that can't explain its own failure modes isn't production-ready. The evaluation layer explicitly documents:
- ·No ground truth labels exist, so precision and recall cannot be computed.
- ·The statistical method assumes approximate normality; bet size distributions are right-skewed, which inflates false positives on legitimate large trades.
- ·No velocity detection means coordinated small-bet patterns evade the system entirely.
- ·Thresholds are empirically unvalidated: a known gap requiring labeled data to close.
This documentation isn't weakness; it's the difference between a demo and a system someone can make decisions with.
Technical stack
- Python
- SQLite
- Pandas
- NumPy
- Next.js
- TypeScript
- Tailwind
- Vercel
- GitHub Actions
- Polymarket REST API
What this demonstrates
The ability to scope an AI/ML problem correctly, build a working detection system on real data, design an evaluation framework that's honest about its own limits, and communicate the gap between prototype and production clearly, without overstating what exists.
That last part is harder than it sounds. Most ML demos claim more than they've proven. This one doesn't.
Selected experience
Production AI and data products shipped under enterprise and regulatory constraints.
Principal Product Manager
Abaxx Technologies
Enterprise communications platform inside a regulated futures exchange.
Compliance and audit integration. Cross-functional product leadership across engineering, design, and operations.
Director of Digital Product
Base Carbon
Carbon market financier.
Scoped and built an AI-driven anomaly detection proof of concept for fiduciary-grade workflows. Authored the Python implementation end to end. Multi-jurisdiction regulatory framework experience.
Senior Product Manager
Hudson MX
Enterprise analytics for the media and advertising industry.
Process design, intake and prioritization, and scaling a product organization over five years.
Full work history on LinkedIn. Additional detail available on intro calls.
How we work
Intake call (30 min)
We talk through what you're shipping, what's at risk, and whether Foundry is the right fit. If we're not, I'll tell you and point you somewhere better.
Scoped engagement
No retainer-for-retainer's-sake. Each engagement has a defined start, defined deliverables, and a defined out.
Senior practitioner, not a junior team
Foundry is led by a senior PM with 10+ years in enterprise product and several years focused on AI in regulated environments. No account managers, no delivery layer, no offshore handoffs.
Honest about constraints
AI is not magic and senior practitioners don't pretend it is. You'll get calibrated estimates and direct answers about what's possible, what's risky, and what's not worth doing.
About
Foundry AI Ops is led by Nicole Myers, a senior product practitioner with 10+ years building enterprise products and the last three years focused on AI product work in regulated environments.
Most recent work: LLM-based extraction and validation systems where output quality, auditability, and compliance review are non-negotiable. Earlier: Principal Product Manager at Abaxx Technologies, Director of Digital Product at Base Carbon, and senior product roles at Hudson MX.
Columbia MS Applied Analytics. Atlanta-based. Daily user of Claude, Cursor, and Claude Code; writes Python when the project needs it.
What I bring
- ·Senior product judgment, not junior delivery
- ·LLM extraction, validation, and eval framework design
- ·Regulated and audit-ready product workflows
- ·0-to-1 product builds and seed-round narrative
- ·Calibrated honesty about what AI can and cannot do
Working on AI inside a regulated,
enterprise, or high-stakes context?
Let's talk.
I read every message personally. Reply within 24 hours.