I built the wrong tool for my sister

Three hundred thousand Chilean pesos.

That's what my sister posted on Instagram on Friday, March 13th, looking for someone to help her for two weeks with a job she described as endless. I saw her story at 12:45 PM and replied right away. That same afternoon we met at a cousin's birthday party and I asked how the search was going. She said they'd found someone who'd start on Monday. And then, as always happens at family gatherings, the conversation drifted somewhere unplanned.

One hundred and two volumes

Silvana studies law and works at a legal firm. She told me they were handling a large criminal case — a prosecution over the misuse of reserved funds in Chile's national police force, with dozens of defendants and years of accumulated investigation. Her firm represents one of them. And Silvana's problem wasn't the case itself, but the mountain of paper that came with it.

The case's investigative file contains one hundred and two volumes. One hundred and two PDF files, each between three hundred and eight hundred pages long, filled with documents that piled up with no chronological or thematic order. Witness statements, official letters, bank records, expert reports, photocopies, scans — everything mixed together, everything scattered across thousands upon thousands of pages.

Her job was to cross-reference those one hundred and two volumes with a prosecution document to fill in a spreadsheet with specific data for each document: volume number, page range, evidence number in the prosecution, type of evidence, date, a brief content description, whether it related to the defendant they represent, who signed the document, and who received it.

For every row in that spreadsheet, the process was the same: open a volume, identify a document, go to the prosecution file, search whether that document appeared as testimonial, documentary, expert, or other type of evidence, note the corresponding number, go back to the volume, extract the data, fill in each column. Next document. Next volume. One hundred and two volumes.

Three hundred thousand pesos for two weeks of manual help to do exactly that.

Forty-eight hours

As she was telling me this, I was already writing code in my head. Ten years of building software does that to you — you hear a manual process and automatically start designing the solution. That night, when I got home, I had the architecture roughly mapped out. By Sunday, March 15th — two days after the party — I had a working prototype deployed. I sent her a link to try it.

Forty-eight hours from birthday party to deploy. Upload a PDF, get back an Excel file with the extracted data.

I felt proud.

But what Silvana needed wasn't extracting data from a single PDF. It was cross-referencing information across dozens of documents, interpreting the legal context of each one, identifying whether a document corresponded to testimonial or documentary evidence within the prosecution file, finding sequential numbers the prosecutor had assigned following the order of the volumes, and determining whether each document had any connection — any connection at all — to the defendant they represent. A mention, a signature, a name on a list.

My prototype did something useful in the abstract, but it didn't solve her real problem. I had built a solution for the problem I understood between empanadas and drinks at a birthday party, not for the problem she dealt with every day at her desk.

I asked her to explain everything in detail so I could get it right. And she trusted me without hesitation.

Fifteen minutes

On Tuesday, March 17th, she sent me three things on WhatsApp: a fifteen-minute voice note, a two-minute one, and a seven-minute video. Twenty-four minutes of content. She explained the entire case from the beginning — how the investigation originated, why the file was so massive, how the volumes were organized, what each spreadsheet column meant, how testimonial evidence differed from documentary evidence, why accuracy in prosecution numbers was critical, and what each signature and receipt field represented.

I watched the video. But the voice notes — I looked at them and thought: this is too much to process by ear.

Fifteen minutes of WhatsApp audio doesn't feel like fifteen minutes. It feels like an hour when you're trying to take notes. Pause, write, rewind, listen again, realize you missed a detail three minutes back. And in this case, every detail mattered. If I confused a type of evidence, if I misunderstood how documents were cross-referenced with the prosecution file, if I didn't grasp the difference between a signed document and a received one, I'd end up building the wrong tool a second time.

I couldn't afford that.

Sooo slowww

From March 17th I didn't touch the project. Not because I didn't care, but because I knew that building without truly understanding was exactly the mistake I'd just made. And Silvana had her assistant. The work would move forward.

Right?

On Sunday, March 29th, at 10:39 PM, I sent her a message: "How's it going with the case?"

Six minutes later: "Sooo slowww."

She replied instantly. On a Sunday night at ten, in six minutes, because the problem was still right there, on top of her. Nearly two weeks after hiring help, with a person dedicated exclusively to this, and they were still crawling. Because the work is inherently slow when it's manual. It doesn't matter how many hands you throw at it — if the process is opening a five-hundred-page PDF, searching for a name, cross-referencing data with another document, and filling in a spreadsheet cell by cell, it's going to be slow. Always.

That night I sat down in my home office, in the back room of my apartment in Puente Alto, and opened the conversation with Silvana's voice notes. But I didn't open the code editor. Not yet. I opened the audio messages and thought: if I'm going to do this right, I first need to capture every word she said. No assumptions. No interpreting by ear. No missing details.

And that's when the question hit me: why am I about to spend forty minutes transcribing audio by hand when I can build something to do it in thirty seconds?

Voxcribe

That's how Voxcribe was born.

Voxcribe takes an audio file, transcribes it using artificial intelligence, analyzes the content to generate a title and context summary, and delivers a professional PDF ready to download. You upload the audio, wait a few seconds, and get a document with all the information structured — ready to read, share, or feed to an AI to help you properly define the problem.

Which is exactly what I need to do. Transcribe Silvana's voice notes, have the complete transcription in black and white, and sit down to design the right solution. This time without assuming anything. This time with the exact requirements.

And now I have exactly that.

At 2:15 AM on March 30th, I pushed the first version. Minutes later I tested it with Silvana's actual voice notes — the same ones she'd sent on March 17th, sitting in my WhatsApp for nearly two weeks. The fifteen-minute audio became a four-page PDF with the complete explanation of the case, the structure of the one hundred and two volumes, the types of evidence, and every detail I need to build the tool that will actually serve her. The two-minute note captured a clarification about testimonial evidence that would have been easy to miss while listening by hand.

At 2:21 AM, the MVP was validated. With the real audio. With the real case. With the certainty that not a single word was lost.

Voxcribe wasn't the project I wanted to build. I wanted to build the tool that solves Silvana's problem — that comes next, and it's called Herlax. But to build it right, I first needed to truly listen. And to truly listen to a fifteen-minute voice note without missing a single detail, I needed Voxcribe.

I built it for myself. But anyone who receives long voice notes — meetings, instructions, interviews, client feedback — has the same problem. I needed to understand my sister. Someone else needs to understand their boss, their client, their team.

Because "sooo slowww" isn't a phrase I want to hear again. And three hundred thousand pesos for two weeks of manual labor is a price nobody should have to pay when technology can do it better.

Try Voxcribe for freeVoxcribe is available for free at kainext.cl/tools/voxcribe.

Sometimes the best product you can build isn't the one you had in mind, but the one you need to get there.