Building intelligent
defenses against
human & AI adversaries

# Anti-Fraud Systems

Bootstrapped **3 anti-fraud areas** combining human-in-the-loop
and ML-driven solutions. Drove bottom-line savings with
minimal user friction across Dasher, Consumer, and Merchant.

## Mitigations pioneered
- GPS location spoofing detection
- Mobile app hooking attacks
- Fake identities & synthetic accounts
- Account compromise + account sharing

stack: XGBoost · LightGBM · human-in-the-loop review queues

# Transformers & LLMs

Ship production transformer systems, from encoder-based classifiers
to agentic LLM applications. Focus on fine-tuning, evaluation, and
measurable outcomes.

## Models worked with
- **Encoders**: BERT, DeBERTa
- **Embedding models**: SBERT, Harrier-OSS
- **Open-weight generative**: Llama, Mistral, LLaVA, Phi, Qwen, DeepSeek, Mixtral, Gemma, Kimi
- **Proprietary**: GPT, Claude, Gemini
- **Reasoning**: o-series, Claude extended thinking, DeepSeek R1, Gemini thinking

## Applications
- **Agentic AI**: multi-step reasoning with tool use, MCP, and chain-of-thought
- **Fine-tuning** BERT-family encoders on adversarial samples
- Agentic pipelines that identify classifier blind spots and curate targeted training data
- **ASA**: autonomous threat investigation pipeline
- RAG + semantic search over security corpora
- Multimodal prompting for classifier development

> Strong bias toward production-grade deployments with clear KPIs

# Classical ML & Product Ownership

Product owner across the classical ML stack, intimately involved
in model design, training data strategy, evaluation, and deployment.
The unglamorous models that quietly keep fraud rates low and
pipelines running are the ones I tend to own end-to-end.

## Models & techniques
- **XGBoost / LightGBM**: fraud scoring, risk classification, threshold tuning
- **OCR & QR scanning**: document extraction and barcode pipelines
- **SHAP**: model explainability for trust, review queues, and compliance
- Anomaly detection: unsupervised + semi-supervised methods
- Bayesian inference for uncertainty quantification in scoring
- Precision / recall tradeoffs anchored to business cost functions

## Process
- Balance speed vs. accuracy for real-time scoring systems
- A/B test design, experiment analysis, and metrics definition

stack: XGBoost · LightGBM · scikit-learn · SHAP · PyTorch

> The load-bearing workhorses, not the showhorses

# Engineering Depth

Shipped production backend systems across enterprise and startup
environments, from microservices to CI/CD infrastructure.

## Platform
- AWS: ECS, Lambda, S3, RDS, microservices at scale
- Docker: multi-stage builds, containerized deployments
- Linux: system administration, iptables, server hardening

## Practices
- **CI/CD**: reduced test cycle from 1 week → 9 minutes
- REST APIs with JBoss/Wildfly, Hibernate, async processing
- Multithreading, high-performance backend systems

stack: Python · Java · AWS · Docker · Linux · JavaScript

# Detection Platform

Product manager on **Sublime MQL**, the domain-specific query
language behind our continuously-iterated open detection library.
Work with the detection team to identify efficacy gaps and shape
new MQL functionality as attacks and threats evolve.

## Shipping at scale
Work within tight runtime, RAM, and cost budgets across both
cloud and self-hosted deployments. Continuously balance latency,
detection efficacy, cost, and scale.