Hi HN,
My name is Jeff, and I'm the founder of InputRight. For 20 years, I was a contractor. The constant anxiety wasn't the work itself, but finding the next lead to pay the bills. That pain drove me to transition from landscaping and exterior work into digital marketing and lead generation.
For the past 15 years, I've focused on helping small businesses get new clients, and for the last four, my work has centered on leveraging AI for digital marketing/SEO/leadgen, as well as designing an AI Co-pilot for small businesses.
Last month, I was launching a new lead-gen service for contractors and realized a critical missing piece: a voice agent on their websites to capture leads 24/7. But there was a huge problem—I couldn't find a single voice bot that could guarantee accuracy. For a contractor, where a single lead can be worth tens of thousands of pounds, a misheard digit in a phone number is a disaster.
I knew this was a problem that needed solving. And it wasn't just for contractors or lead generation. I quickly realized that any business where data accuracy is paramount—like patient intake, customer support, or financial applications—could benefit from this technology.
I'm not a developer, so I took a unique approach, using a top-down workflow with LLMs to build this platform. My process looked something like this:
High-level architecture: I used a custom-prompted LLM in Google AI Studio to act as my "CTO" and outline the project's entire structure.
Code generation: The high-level plan was fed into another LLM, my "Senior Software Engineer," who provided specific commands and code snippets.
Execution: I used the Gemini CLI as my "junior developer," running the commands.
I'd then copy the outputs, including any errors, back into the AI "engineer" for advice on what to do next. The LLM would often instruct me to go to websites and scrape pages when it needed additional context. This entire process was a whirlwind of frustration and creativity. Debugging without being able to code was a huge challenge. I had to rely on the LLM to analyze error messages and explain what was wrong, which often felt like translating a foreign language.
An interesting observation was that once I got to around 600,000 to 700,000 tokens in my Gemini 2.5 Pro context, the LLM's performance would begin to degrade. It was then necessary to instruct the LLM to create a clone of itself by generating a prompt that would start a new project for a new AI engineer to continue the job.
The result is InputRight, an open-source voice agent that captures data with 100% accuracy. The key is a simple, human-in-the-loop verification step: the bot transcribes the user's input, displays it in a clean form, and asks the user to confirm or edit via voice or manually. This bridges the gap between the speed of voice and the certainty of a traditional form.
It's surreal to me that in less than four weeks, I've gone from an idea to a working, open-source MVP on GitHub. I'm sure there are architectural mistakes and the code commenting may be poor, but the fact that it exists and works is mind-blowing.
I'd love to hear your feedback on the project itself and my process.
Last month, I was launching a new lead-gen service for contractors and realized a critical missing piece: a voice agent on their websites to capture leads 24/7. But there was a huge problem—I couldn't find a single voice bot that could guarantee accuracy. For a contractor, where a single lead can be worth tens of thousands of pounds, a misheard digit in a phone number is a disaster.
I knew this was a problem that needed solving. And it wasn't just for contractors or lead generation. I quickly realized that any business where data accuracy is paramount—like patient intake, customer support, or financial applications—could benefit from this technology.
I'm not a developer, so I took a unique approach, using a top-down workflow with LLMs to build this platform. My process looked something like this:
High-level architecture: I used a custom-prompted LLM in Google AI Studio to act as my "CTO" and outline the project's entire structure.
Code generation: The high-level plan was fed into another LLM, my "Senior Software Engineer," who provided specific commands and code snippets.
Execution: I used the Gemini CLI as my "junior developer," running the commands.
I'd then copy the outputs, including any errors, back into the AI "engineer" for advice on what to do next. The LLM would often instruct me to go to websites and scrape pages when it needed additional context. This entire process was a whirlwind of frustration and creativity. Debugging without being able to code was a huge challenge. I had to rely on the LLM to analyze error messages and explain what was wrong, which often felt like translating a foreign language. An interesting observation was that once I got to around 600,000 to 700,000 tokens in my Gemini 2.5 Pro context, the LLM's performance would begin to degrade. It was then necessary to instruct the LLM to create a clone of itself by generating a prompt that would start a new project for a new AI engineer to continue the job.
The result is InputRight, an open-source voice agent that captures data with 100% accuracy. The key is a simple, human-in-the-loop verification step: the bot transcribes the user's input, displays it in a clean form, and asks the user to confirm or edit via voice or manually. This bridges the gap between the speed of voice and the certainty of a traditional form.
It's surreal to me that in less than four weeks, I've gone from an idea to a working, open-source MVP on GitHub. I'm sure there are architectural mistakes and the code commenting may be poor, but the fact that it exists and works is mind-blowing.
I'd love to hear your feedback on the project itself and my process.