Case Studies

Find the real bottleneck. Eliminate it. Ship results.

Every case study on this page follows the same pattern: the problem everyone sees isn't the real problem. The real problem is the interface between human understanding and machine capability. Find the actual bottleneck. Eliminate it. Everything else follows.

FEATUREDCOGNITIVE PROSTHETICOPEN SOURCE

Brain MCP — A Cognitive Prosthetic for One Person

The Context

ADHD means monotropic attention. One tunnel at a time. Total immersion — and total amnesia for everything outside it.

Every context switch was a factory reset. Decisions remade. Questions re-asked. Breakthroughs rediscovered months later. Not poor organization — neurology.

The Problem Everyone Was Solving

Every AI memory tool is built for organizations. Team knowledge bases. Corporate wikis. Institutional memory for groups of people. The assumption: memory is a collaboration problem.

But the bottleneck wasn't AI capability. It was context continuity for a single mind. Three months deep in Torah study? The AI architecture decisions don't exist anymore. Switch to frontend? The 47 open Torah questions vanish. Not deprioritized. Gone.

The Real Bottleneck

“The prosthetic doesn't fix the brain — it works around it. Don't fight monotropic attention. Build infrastructure that preserves context across attention shifts.”

Every AI memory tool is built for organizations. Brain MCP is built for how one person actually thinks.

The Solution: 25 MCP Tools

387K messages indexed across Claude, ChatGPT, Claude Code, and Clawdbot. 85K semantic embeddings for conceptual search. 9,979 structured summaries with extracted decisions, questions, and breakthroughs. Running cost: $0.05/day.

The trigger phrase: use brain — and 12ms later, the full context is back.

tunnel_state
"use brain" — reconstruct save-state for any cognitive domain in 12ms.
context_recovery
Full re-entry brief: thinking stage, open questions, last decisions.
switching_cost
Quantified cost of moving attention between domains. Makes the invisible visible.
trust_dashboard
System-wide proof the safety net works. Coverage, sync health, freshness.

+ 4 more prosthetic tools (open_threads, dormant_contexts, cognitive_patterns, tunnel_history) and 17 generic tools (semantic search, thinking trajectory, alignment check, etc.)

The Results

MetricBeforeWith Brain MCP
Context recoveryGone when tunnel moves12ms
Open questionsForgotten permanently111,942 preserved
DecisionsRemade from scratch36,743 tracked
Domain switchingFull restartCost quantified + brief
Running cost$0.05/day
Context survival0%100%

One person. 18 months. A cognitive prosthetic that turns neurological context death into queryable memory.

Now open source — MIT licensed. Because this problem isn't unique to one person.

Timeline

2023-06First AI conversations begin(Raw data accumulates)
2024-08Context death problem crystallizes(Pattern recognized)
2025-02Parquet pipeline built(Messages become queryable)
2025-06LanceDB vector search added(Semantic search enabled)
2025-11Cognitive prosthetic tools built(8 tools for monotropic attention)
2026-02Open source launch — MIT licensed(Production-grade, 143/143 tests)

Technical Stack

LanceDB + DuckDB
Vectors + Structured data
nomic-embed-text-v1.5
768d embeddings, local
Model Context Protocol
25 tools, any client

Same Pattern, Different Domains

Accessibility

QinBot

Problem: Religious communities need kosher phones (no browser, no apps). But they also need AI.

Real bottleneck: Everyone assumed AI requires a smartphone. The real constraint was the interface assumption, not the hardware.

Result: Full Claude AI on a dumb phone. No browser, no apps, no images. The most capable kosher phone on the planet.

Zero apps • Full AI • Kosher phone
Data Engineering

YouTube Pipeline

Problem: Everyone thought the bottleneck was transcription cost — API fees, cloud compute, scaling.

Real bottleneck: The real constraint was dependency on external services. Remove the dependency, remove the cost.

Result: 32K videos transcribed. 41.8M words processed. Local ML on Apple Silicon. Total cost: $0.

32K videos • $0 cost • Local ML
Automation

Google Apps Script Portfolio

Problem: 41 separate business processes running on manual labor and email chains.

Real bottleneck: The constraint wasn't technical complexity — it was that nobody realized these were automatable. The bottleneck was imagination.

Result: 41 production scripts across 13 categories. SSN matching, Torah translation, WooCommerce automation. All running unattended.

41 scripts • 13 categories • Zero maintenance
View all repos on GitHub

Have a bottleneck that needs finding?

48-hour diagnostic. Premium hourly. Async-first. Steve qualifies, I engage.

How I Work