Building Tether — From Noisy

Building Tether — From Noisy Telemetry to Deterministic Operations

Role: Lead Architect & Full-Stack Developer Tech Stack: React, Cloudflare (Pages, Workers, R2, Zero Trust), Google Cloud Platform (Cloud Run, Cloud Storage), Python, FastAPI, Scikit-Learn.

The Challenge: The Hospitality Data Gap

Modern hospitality operators are drowning in data but starving for actionable intelligence. A restaurant's two most critical systems—the Point of Sale (revenue) and the scheduling platform (labor)—operate in complete isolation. Because these systems do not dynamically communicate, managers are forced to make high-stakes labor cuts on the fly based on delayed reporting and gut feeling.

This disconnect results in thousands of dollars of weekly margin bleed. The challenge was clear: build a system that bridges these fragmented APIs, normalizes the data, and provides real-time operational certainty.

The Solution & The Product Pivot

I engineered Tether to be an AI-native operational layer for restaurant management. However, the true breakthrough of this project wasn't just technical—it was architectural.

Initially, I designed Tether as a "Live Data Prediction Tool" that used active telemetry to drive real-time floor decisions. Through testing and auditing the data streams, I identified a critical UX flaw: live data is inherently noisy and reactive. To solve this, I executed a complete priority inversion, refactoring the application state to a "Schedule-First" philosophy.

Instead of chasing live data, Tether now ingests historical data to generate a deterministic, highly optimized 14-day schedule baseline. The machine learning models were strategically demoted from "decision makers" to "real-time guardrails." Once the floor opens, Tether acts as a safety net, validating execution against the baseline and alerting managers to profit leaks before they compound.

Technical Execution: A Masterclass in Edge ML

To ensure security, scale, and sub-100ms latency, I architected Tether as a zero-backend Single Page Application (SPA) driven by serverless microservices.

Edge Infrastructure & Security: The frontend is deployed via Cloudflare Pages and secured behind a Cloudflare Zero Trust perimeter, requiring One-Time PIN (OTP) authentication for operator access.

Data Normalization: I developed Cloudflare Worker proxies to securely handle OAuth handshakes, ingest data from POS systems (Square, Toast) and labor platforms (7shifts), and normalize the varied streams into a unified, sanitized client schema.

Autonomous ML Pipeline: I engineered a fully autonomous, serverless retraining loop hosted on Google Cloud Run. Every Tuesday at 3:00 AM UTC, the pipeline wakes up, pulls historical telemetry from Cloudflare R2, and retrains the primary Approval and Labor-to-Sales (LTS) models (using Ridge and Logistic Regression).

Strict Data Contracts: The ML pipeline strictly enforces a 63-feature data contract. It validates baseline accuracy and ensures zero NaNs before allowing any model to pass into production, guaranteeing operational stability.

Highly Optimized Model Distribution: Fresh model weights are served to the browser via a Dockerized FastAPI microservice (kept aggressively lean at ~500MB) and distributed globally through Google Cloud Storage (GCS).

The Business Impact

Tether replaces the anxiety of restaurant management with mathematical certainty.

By automating the schedule generation and monitoring real-time Labor-to-Sales (LTS) velocity, Tether catches margin bleed live—such as a sudden drop in patio sales pace due to weather. It translates complex ML predictions into simple, actionable alerts (e.g., "Trim one support role. Protects $190 margin.").

The result is protected daily profit margins, guaranteed labor compliance, and management teams empowered to run their floors with absolute confidence.