HITL Sentence Generator: Product Design & Workflow Strategy

Scaling AI data pipelines by replacing fragmented manual processes with a centralized syntax engine.

Case Study: The Modular Utterance Generator

Product Design & Experience Architecture

Project Overview

The Modular Utterance Generator is a visual syntax system designed for high-scale NLP training data. I transformed a fragmented, manual freelancer workflow into a single-screen Human-in-the-Loop (HITL) interface. This enabled linguists and non-technical staff to generate thousands of high-quality training utterances with zero duplicates and 75% faster velocity.

The Challenge: The Scalability Wall

Before this system, e2f relied on manual writing for AI training sets. In a pre-LLM world, variation relied entirely on human memory, leading to a mechanical-sounding data bottleneck:

• Inconsistent Quality: High variance in sentence structures across different writers.

• The Expertise Gap: Massive friction between PhD linguists and entry-level annotators.

• QA Fail-States: 4-week lead times and expensive cycles due to human-generated duplicates.

The Strategic Pivot: 6 Stages of Evolution

I didn't just build a tool; I led the operational shift of the data pipeline:

1. Identification: Pinpointed the manual freelancer system as a non-scalable bottleneck.

2. Validation: Proved a rule-based generator was technically feasible through linguistic analysis.

3. Architecture: Designed the underlying syntax system and "grammar rules."

4. Proof of Concept: Built a functional POC to demonstrate logic-driven generation.

5. Engineering Alignment: Coordinated with developers to move the POC into production.

6. Iterative Redesign: Led a full UX overhaul based on real-world "battlefield" data.

The Solution: "Structure Felt, Not Taught"

I moved away from academic jargon toward a spatial, component-based UI. The system anchors the entire workflow on a single surface to reduce context switching and mental fatigue, organized into four layers:

1. The Structural Layer (Define the Pattern)

Users build patterns (e.g., Verb → Object → Location) using scrollable columns.

• Design Choice: I stripped linguistic jargon (e.g., "Noun Phrases") in favor of functional labels, reducing training time.

2. The Content Layer (Populate Wordlists)

Structural slots expand into flexible wordlists (e.g., "Play," "Start," "Shuffle").

• Design Choice: The UI avoids auto-correct to ensure data doesn't feel mechanically assembled, preserving natural human variation.

3. The Feedback Layer (Real-Time Preview)

A live preview updates instantly as users manipulate structure or content.

• Design Choice: This immediacy lowers cognitive load by making cause-and-effect transparent.

4. The Delivery Layer (Export)

Data is exported pre-formatted for engineering pipelines.

• Design Choice: The UI prevents duplicates at the point of creation rather than fixing them in post-production.

Key Outcomes

• 75% Velocity Increase: Reduced data generation timelines from 4 weeks to 1 week.

• Zero Duplicates: Systematic generation eliminated the #1 QA fail-state.

• Operational Alignment: Created a "Single Source of Truth" used by Engineering, Ops, and Linguistics.

• Enterprise Scalability: Successfully handled complex syntax for Fortune 500 clients.

Reflection & Future State

This project illustrates my ability to simplify complex logic through intuitive interaction. By keeping the workflow on a single surface, I created a system that is accessible for novices while remaining powerful for experts.

The Next Evolution: I am currently exploring a Hybrid Human-AI workflow, using LLMs to suggest wordlist expansions while keeping the human Architect in control of the structural integrity.