Ethan – Pandas, Python, SQL

AI traffic-surveillance company supplying Argentine municipalities on a hardware plus subscription model.

• Served as the sole technical founder of a customer-facing ML product running unattended in production across about 30 cameras and 10 municipalities.
• Designed an edge plus central-AWS inference architecture with an automated normalization loop that self-adjusted camera exposure and contrast for changing light conditions.
• Built a YOLO-based pipeline detecting helmet, over-capacity, and headlight infractions.
• Ran fleet operations through an agentic LLM layer where a tool-calling agent connected to cameras over SSH, ran diagnostics and health checks, and applied fixes across the fleet.
• Navigated multi-year B2G procurement and political turnover to keep the business operating through full electoral cycles.

Real-time P2P conversation moderator. A streaming pipeline of buyer and seller interactions scored for suspicious activity. Reached a peak of over 1k QPS. Daily scammer detection grew by 500%, and over the course of 4-5 months, most of the most active bad actors had left the platform. App reviews of the Paxful platform saw a 25% gain, an unprecedented increase.

• Ran a structured roadmap-discovery process with stakeholder interviews, data audits, shallow prototypes, and impact/feasibility scoring to choose and ship the highest-ROI ML initiative.
• Designed and shipped a Kafka-based real-time scoring pipeline processing about 1M events per month.
• Served XGBoost from a Python/Flask microservice over feature vectors built by streaming Kafka aggregators.
• Balanced automated decisions with human-in-the-loop review through a two-tier action policy with anti-jitter logic.
• Quadrupled daily fraud catch while reducing required human review from 10 to 2 FTE.
• Shadow-launched the system against the existing moderator workflow and tuned policy decisions as capacity scaled.
• Engineered features across account history, fraud-graph neighbors, IP/geo clustering, trade-size deviation, conversation cadence, and TF-IDF tokens.
• Built an agentic case-review system where a tool-calling LLM agent assembled account history, trade records, and fraud-graph context into a single case brief.
• Hardened the LLM layer with an evaluation harness using LLM-as-judge scoring, prompt versioning, and regression tests.
• Automated daily Slack reporting on volume, prediction distribution, and TP/FP segmented by geography.

A non-profit Torah study media library that offers over 1,000 animated videos, podcasts, in-depth courses, and printed guides. Founded by the renowned educator, its mission is to help people discover the beauty, spiritual meaning, and relevance of biblical texts through close literary analysis

• Led the platform migration from Gatsby to Next.js, restoring engineering velocity and reliability across an iOS/Android/web/backend product.
• Reduced time-to-first-frame from about 18 seconds to sub-second on 4G mobile through architectural redesign and targeted performance engineering.
• Drove user growth and revenue gains by replacing the freemium model with a trial-based conversion flow and a redesigned onboarding experience.

Online marketplace for used industrial machinery.Online marketplace for pre-owned industrial machinery.

• Owned the full marketing-and-product ML surface across recommendations, retention, lifecycle messaging, and content categorization.
• Shipped the company's first recommendation engine using TensorFlow on AWS and a suite of ranking algorithms inside an in-house Python ML framework.
• Built embedding-based listing similarity for search, deduplication, and related-machine recommendations.
• Built a personalized email send-time system using time-series analysis that improved open rate and email-attributed revenue.
• Shipped the company's first customer churn model.
• Integrated LLMs using GPT for product categorization at catalog scale with a confidence-gated human-in-the-loop review queue and ongoing evaluation.
• Reduced uncategorized listings from 15% to 2%.
• Worked hands-on in Jupyter, SQL, and Pandas day to day.

A hedge fund that uses quantitative analysis to make investment decisions.

• Built the fund's first end-to-end automated trading system covering data ingestion, feature engineering, model training, and operationalization into live trading flow.
• Modeled equity-market behavior using alternative data sources.

The largest e-commerce platform in Latin America.

• Specialized in fraud prevention.
• Built and deployed ML models for transaction-level fraud scoring that reduced fraudulent transactions by more than 70% across four operating countries.
• Designed custom neural networks in the pre-framework era and implemented them directly from research papers.
• Benchmarked Random Forests, SVMs, and GBMs in R.
• Trained models on AWS Linux infrastructure.