Ph.D. · MBA · CAMS · 13+ years

Hugo Vo

Senior Product Manager: Data, AI/ML & GenAI Products

I take data, machine learning and GenAI/LLM products from discovery to production in regulated financial services, and I stay accountable for adoption, trust and measurable business impact. My focus is AI that model-risk, compliance and business leaders can explain, audit and sign off.

linkedin.com/in/hugovoHouston area, TXDownload PDF

Measured impact

Results from products taken from concept to production.

10×Defect detection rate (~2% → ~20%)ML decision-quality review system · patent pending US 2026/0187638 A1
100%Review coverage, up from a ~5% manual sampleSame system, now the foundation of an enterprise quality product
+30%Detection yieldAutomated rule-tuning framework · patent pending US 2026/0260242 A1
−40%Capability cycle timeSame framework, scaled across two product lines
−60%False-positive volumeStatistical threshold tuning with detection effectiveness preserved
~$500KFraud losses cut in year oneAnti-fraud engine taken from concept to multi-channel deployment

Selected work

Production products, plus independent prototypes built on synthetic data.

Responsible-AI layer for model governance

Production
Problem
Governance reviewers and executives needed to understand complex model-tuning analysis quickly and consistently.
Approach
An LLM turns deterministic statistical results into review critiques and executive summaries. Output is schema-constrained and advisory only; a named human signs off; the AI can be switched off, with a non-AI fallback.
Result
Shipped to production with model-risk governance gates. The AI can never override a deterministic pass/fail decision.

Automated rule-tuning framework

Production
Problem
Tuning detection rules by hand was slow, inconsistent between cycles and hard to defend.
Approach
Co-invented an automated framework that evaluates candidate settings, scores them against a stated risk appetite and produces a governance-ready review package.
Result
+30% detection yield, −40% cycle time. Patent pending; scaled across two product lines and through a company acquisition.

ML decision-quality review

Production
Problem
Quality control relied on reviewing a small random sample of decisions, so most defects went unseen.
Approach
Hybrid NLP and gradient-boosted models with SHAP explanations score every decision and point reviewers at the likeliest defects, reconciled against real outcomes.
Result
100% coverage (from ~5%) and 10× defect detection. Patent pending.

LLM investigation copilot

Independent prototype
Problem
Open-ended LLM answers can't be trusted in an investigation unless every claim can be checked.
Approach
Designed the grounding, citation and system-prompt architecture for an LLM copilot inside a full-stack case management prototype (FastAPI, React/TypeScript, Docker).
Result
Every answer links back to the evidence it came from.

PII protection for unstructured text

Independent prototype
Problem
Free-text narratives are valuable for ML but full of personal data.
Approach
Evaluated 7 architectural options on accuracy and operating cost, then chose a layered, fail-closed design with a pass / review / block release gate.
Result
Benchmarked on a 1,000-record synthetic corpus, with a governed release path.

Champion–challenger & red-team lab

Independent prototype
Problem
New models often reach production without proof that they beat the current one or hold up under attack.
Approach
A shared lab where challengers (graph algorithms, NLP methods, ML models) compete against the incumbent on accuracy, latency and adversarial cases.
Result
Nothing is promoted without passing a gatekeeper, and every result is reproducible.

How I build AI products

  1. Start from the decision, not the model. Define who acts on the output, what success looks like and how adoption will be measured.
  2. Rules first, LLM where it adds value. Deterministic logic handles routine cases; the LLM covers the rest, with cost controls.
  3. The LLM recommends; it never decides. Human sign-off and deterministic checks stay the backstop.
  4. Set kill criteria before the build. Pilots are measured against a stated bar, and a failed bar is a valid answer.
  5. Explainability, fairness, PII controls and auditability are designed in, not retrofitted for model risk review.

GenAI & ML

  • LLM product design
  • Claude API
  • Prompt engineering
  • RAG & grounding
  • Agentic workflows
  • XGBoost
  • SHAP
  • NLP / transformers

Product

  • Discovery → production
  • Roadmaps & PRDs
  • User stories & acceptance criteria
  • Agile / Scrum
  • KPI definition
  • Change management

Platform & data

  • Python
  • SQL
  • FastAPI
  • React / TypeScript
  • Docker
  • REST APIs
  • AWS (working knowledge)

Governance

  • SR 11-7 / OCC model risk
  • PII & data classification
  • Access controls
  • Audit trail design
  • CAMS