x:0 y:0 Speech Therapy / 01
Project · Clinical tool · AI-built

Speech Therapy App

A daily home practice app for a single kid, one target sound, evidence-based methodology. Four AI agents built it. One human orchestrated.

Speech therapy works in the office. The harder problem is carryover, making new speech patterns stick across the other 166 hours of the week. This app runs the evidence-backed carryover protocol at home. That is what it does. The interesting part is how it got here.

Research, product design, coordination, and implementation each had a dedicated AI role, with me reviewing the work. The result is a concrete example of what AI-to-AI product delivery can look like when the research, the spec, and the build each have a clear owner.

Built byRex (research) · Harper (spec) · Cal (coordination) · Draper (build)
EvidenceLLM Wiki · 7 compiled articles · 10 academic sources
ScopeOne child · one target sound · single-kid app
PostureHome practice · parent-led · no cloud
x:840 y:0 The Build / 02
How this actually got built

One human. Four AI agents. A shipped app.

This started as an experiment in seeing what my agents could do together. The work flowed through a repeatable AI build workflow with clear scopes, reviewable handoffs, and human oversight.

AGENT CHAIN / ONE HUMAN, FOUR AGENTS PHASE 1 · RESEARCH OPERATOR Dustin orchestrator REX research agent compiles the wiki LLM WIKI 7 articles 10 academic sources ASHA · RCTs · Marshalla knowledge base by nvk PHASE 2 · BUILD HARPER homeschool agent product side SPEC evidence-grounded spec CAL chief of staff routes the work DRAPER app developer writes the code SHIPPED APP daily tablet app start research compiles wiki evidence feeds Harper writes hands off starts build ships daily use · next iteration LEGEND human AI agent artifact hand off informs / loops N
Research and build run as separate phases, each in the scope of the agent who owns it. What makes this repeatable is not any single model. It is the shared context, source-grounded handoffs, and human review before durable changes.
x:0 y:860 What They Built / 03
The product itself

Five-step session. Eight difficulty levels. Clinical guardrails.

Each daily session follows a fixed protocol. The sequence is not configurable. It reflects the evidence base Harper compiled before a line of code was written. Advancement is locked to clinical thresholds. Escalation triggers fire automatically when patterns suggest the speech-language pathologist needs to weigh in.

SESSION FLOW / FIXED PROTOCOL · 5 TO 10 MINUTES WARMUP ear discrimination PROBE session baseline DRILL main practice · random order EMBEDDED carryover moment SELF-RATE metacognition PROGRESSION / 8 LEVELS · ISOLATION TO CONVERSATION L1 Isolation L2 Syllable L3 Word L4 Phrase L5 Sentence L6 Structured L7 Spontaneous L8 Generalize
Cadence5 to 10 minutes · daily target · 4 to 5x per week minimum
AdvancementLocked thresholds · 2-session minimum · auto drop-back on failure
Escalation5 clinical triggers fire automatically (plateau, regression, avoidance)
Self-monitoring4 progression levels · post-hoc rating through spontaneous self-correction
x:1240 y:860 By The Numbers / 04
What AI-to-AI delivery actually produced

The rigor, measured.

Depth of work that a traditional small-team build would have cut for scope. The agents did it because their scopes were clear and their handoffs were cheap.

7

Research articles

Rex compiled into the LLM Wiki.

10

Academic sources

Preston 2020 (ASHA), Sjolie 2016, Raaz 2025 RCT, Pritchard 2025, Marshalla, Kim 2025 ASR, ASHA Leader 2018, plus three more.

8

Difficulty levels

Isolation through generalization.

5

Escalation triggers

Auto-fire to flag the clinician.

4

Self-monitoring tiers

Carryover keystone, per Marshalla.

0

Human code lines

Written by me. The four agents owned the build.

x:0 y:1620 Stack & Use / 05
Under the hood · and why it stays small

Boring stack. Explicit test surface. No cloud.

Stack

Vite, React, TypeScript. No backend.

Vite + React + TypeScript. Pure browser localStorage, no database, no accounts. The goal was to keep the first version as simple as possible: enough structure for the agents to build together, and enough product to test with my child at home.

It can grow if it needs to: add a backend, accounts, richer reporting, or a more formal content workflow later. For now, the important thing is a useful local app with clear clinical rules, tested progression logic, and a small surface area.

Status

In active daily use.

Core drill engine and session flow are complete. Daily sessions are running at home. Audio synthesis is in validation. A test set of clips is being evaluated before the full batch is generated.

PlatformTablet · web app · local network
StageProduction, daily use
Session5 to 10 minutes

Use

Private household tool.

This is not a public product or a shared app. It is a focused household tool built to support one child's daily practice and to learn how my agents collaborate across research, specification, coordination, and implementation.

next dragpan 0fit all Eschome
Want to reach out? X is best.
x / @dpersek linkedin github / syntaxsawdust