AI Platform @ FenixCommerce

Project Timeline: 2025 – present

Skills: Python, XGBoost, LightGBM, PyTorch, Polars, MLflow, AWS Athena, HubSpot API, Amazon SES, Pydantic AI, Claude API, SQLite, Telegram Bot API

Overview

Built and operates an autonomous AI agent platform (“agentvault”) at FenixCommerce that coordinates specialized AI agents across four production systems. 30+ agent sessions completed with zero data loss.

Systems

EDD Prediction (fenixlearn)

Replaced FenixCommerce’s production delivery date prediction system. Improved on-time rate from 41.8% to 67.0% (+25% on-time improvement, +60% relative) across 5.96M orders using XGBoost with ordinal classification and asymmetric loss. Validated on 78 walk-forward temporal folds. Proved 80%+ on-time feasible with real-time scan events (+13pp additional in PoC). 326 segments probed across 10 dimensions, 15 model architectures tested.

A/B Test Incrementality (incremsim)

Measures where the EDD widget helps vs hurts revenue, then optimizes placement. Baseline measurement: $3.46M incremental revenue across 7 retailers (1.88% lift on $183.4M base). Zone/state optimization added $1.48M on top by turning OFF EDD in segments where it hurts conversion, bringing the total to $4.93M (+43% over unoptimized). Per-retailer examples: KUIU went from $443K to $1.49M with 32 state flips (3.4x), Rylee+Cru went from net-negative (-$60K) to +$121K after 44 flips. Automated Mon+Wed pipeline with MLflow logging, heatmaps, and PDF reports.

AI Sales Operations (salesmon)

Automated cold deal detection and outreach management for the sales team. Classifies 306 HubSpot deals via LLM, surfaces a prioritized 15-deal daily digest to VP Sales, and manages follow-up cadence (bump/reiterate/close stages). Smart client filtering via domain matching + Jaro-Winkler + LLM tiebreaker. 22-column portfolio view with outreach timeline DAG.

Delivery Performance Monitoring (fenixalerts)

Automated daily alerting across all fulfillment centers and key retailers. Detects dark warehouses and week-over-week performance drops, generates AI-ranked business insights (Opus 4.6), and delivers R/Y/G card-based HTML email digests. Replaces manual monitoring with a single daily 6 AM PT scheduler.

Platform (agentvault)

The coordination layer: Telegram command and control, institutional memory, PID-polling queues for sequential agent deployment, MLflow backfill watchers, and structured protocols for deployment, reporting, and session handoff. 709 tests across all repos, 73.5K lines of code, 56 agent changelogs.