Hakivo

01

The Problem

Congress introduces more than 10,000 bills a year. Members of Congress do not read them either, and they are not expected to: they have legislative aides who track, summarise, analyse and flag what matters. Everyone else is expected to figure it out alone, across fragmented government websites, in language written to be litigated rather than read, around a full time job. The gap is not information, because Congress.gov is public. The gap is that the information is only usable by someone who has a staffer.

02

The AI Design Decision

I gave the work to four different AI models instead of one, because the four things this product does have nothing in common. Reading a bill closely is slow, careful work over a long document, and I'm happy to wait a minute for it. The trivia shown on screen while you wait only has to arrive before you get bored. Pulling in the news needs something that searches the live web. Reading the briefing out loud is a different job again. One model doing all four would be worse at three of them.

The voice started on one provider and moved to another halfway through, and the honest reason is that I ran out of credits during the hackathon. The reason it stayed moved is that the new one was cheaper and steadier at the volume a daily briefing for every subscriber actually means. A voice you generate once for a demo and a voice you generate every morning for everybody are different products.

What I did not build is a summariser. That was the obvious version and I tried it. Shortening a bill gives you something you can read and can't use, because the part it cuts is the mechanism, and the mechanism is what decides whether the bill helps you or not. So the analysis is broken into fixed parts instead: what it does, who it affects, what it costs, what could go wrong.

03

The Architecture

1. 01Pull every new bill from Congress and all fifty states
2. 02put them in a queue so nothing waits on a slow government website
3. 03four different AI models each do the one job they are best at
4. 04the finished briefing is written and read aloud
5. 05it is waiting for you in the morning, and you never asked for it

04

The Responsible AI Constraint

A product that tells you what's in a bill is one bad sentence away from telling you how to feel about it, and you'd never catch it. Nobody reads a plain language summary of a law with a lawyer sitting next to them. Whatever the model chose to emphasise is simply what you now believe the bill says.

So the answer isn't a rule the AI is asked to follow. It's the shape of the place the answer gets stored. Every bill has two boxes: the case for it and the case against it. Both have to be filled in. A one sided analysis has nowhere to go. It isn't caught in review, it doesn't fit.

Who wins, who loses and who is affected are three more boxes, for the same reason. Each argument has to carry the evidence behind it, so a claim with nothing under it is incomplete rather than persuasive.

And the bill page puts both boxes side by side, so there's no way to read a bill here and see only one half of it.

05

The Result

Winner of the Grand Prize, number one best overall, at The AI Champion Ship. Live at hakivo.com. It covers every bill in the current Congress plus all fifty states, with 7,262 state legislators loaded across 49 states, each with a photo, and you can find yours by address. It charges $12 a month, with a free tier of three briefings, three tracked bills and three followed members. And the 100 Laws That Shaped America podcast writes itself, records itself and publishes itself to Spreaker with nobody in the loop.

06

If you remember one thing

A one sided analysis of a bill has nowhere to be saved. Every bill gets two boxes, the case for it and the case against it, and you can't leave either one empty. That's a storage decision, not an instruction to the AI.