Parallel-agent AI prompt automation platform
Echo Prompter
An enterprise platform that runs prompts 10–100× faster with parallel agents, and regression-tests AI output for drift and hallucinations.
My role: Architect and developer: .NET core, Python orchestration, testing and observability
- Agents
- .NET
- Python
- ChatGPT · Gemini · Claude
- LLM testing
Overview
Echo Prompter is an enterprise-grade platform for data-driven AI execution, built for a client in the United States. I was architect and developer: a high-performance .NET core that manages parallel execution agents, Python workflows that orchestrate ChatGPT, Gemini and Claude, and testing tools that make probabilistic output measurable.
The challenge
Teams using LLMs at scale hit two walls. Running thousands of prompts one after another is slow, and every change to a prompt or model can quietly change the output in ways nobody notices until a customer does.
The client, a US company, needed a platform that could run large prompt workloads in parallel across ChatGPT, Gemini and Claude, and treat AI output like any other software: tested, measured, and costed.
How it works
- 1Prompt dataset
- 2.NET scheduler
- 3Parallel agents
- 4ChatGPT / Gemini / Claude
- 5Regression checks
- 6Results + cost report
Architecture
The core is a high-performance .NET service that manages a pool of execution agents. It schedules work across agents in parallel, tracks each job's state and exposes everything through a REST API. Moving from sequential calls to parallel agents is what delivers the 10–100× speed-up in prompt processing.
LLM orchestration lives in Python workflows, where the model ecosystem moves fastest. These workflows route prompts to ChatGPT, Gemini or Claude and normalise their responses, so the .NET layer does not need to change when a provider or model does.
On top sits a testing layer built for probabilistic output. Regression suites compare new results against known-good behaviour while tolerating harmless variation, and flag real drift or hallucinations. Load tests show how the system behaves at volume, and real-time cost tracking and observability give each run a known price and a trace.
What I built
Parallel execution agents
A .NET backend schedules and manages many agents at once, taking prompt processing from sequential to 10–100× faster.
Multi-model orchestration
Python workflows route prompts across ChatGPT, Gemini and Claude through one REST API.
Regression testing for AI
Automated regression and load tests designed for probabilistic output, catching drift and hallucinations before they reach users.
Cost and observability
Real-time cost tracking and observability for large-scale runs, so every batch has a known price and a trace.
Engineering decisions
- Split the work by strength: .NET for high-throughput concurrency and agent management, Python for the fast-moving LLM ecosystem.
- Treat AI output like any other code under test: regression suites that tolerate probabilistic variation but flag real drift.
Results
- 10–100× faster prompt execution by running agents in parallel instead of sequentially.
- Automated regression and load testing to detect drift and hallucinations before release.
- Real-time cost tracking and observability for large-scale runs.
- Client feedback: "His contribution to our product has been and will continue to be immeasurable." — Randy Hayes, Loadzilla
Stack
- Backend
- .NET (C#), REST API
- AI workflows
- Python
- Models
- ChatGPT, Gemini, Claude
- Quality
- Regression and load testing for LLM output, cost tracking, observability
Need something like this?
I can build a version of this for your product, your data and your stack.
Start a project