SEC Filings Instant Analysis

Some people want SEC filings really fast. Many of them also want a LLM analysis of the filing.

Forthcoming Product

LLM summaries of SEC 8-K filings within ~1s of a filing being published. $20/month. ETA 1-2 weeks.

Likely Product After That

Bring your own prompt and model for any SEC filing. Pricing tbd.

Use case

A user recently emailed me with their automated workflow:

1. Get notification of new SEC filing 2. Get filing 3. Send it to Claude for analysis 4. Trade based on analysis

And expressed interest in a managed solution. Which, I'm happy to do, because this is a fun problem.

How the new product likely will work (probably)

First, we have to detect the SEC filings. My existing notification systems can detect filings within ~200ms of a SEC filing being published. This is easy to improve on, I think I can get to ~10ms without issue.

Second, we can also anticipate SEC filings. This requires using an algorithm to detect SEC filings faster than most methods people use. For example, I've met with trading firms who had trouble cracking this problem. So, the method despite being open sourced is still niche. I think we can get to ~50% coverage here. That means for an 8-K, we would detect it 2s on average faster than traditional methods.

Third, we have to get the SEC filing. This is a bit tricky, because (2) can actually detect filings before their content is published. So, I'm now going to introduce the engineering portion.

We're going to use a Google Worker Pool. The pool will handle detecting new filings, anticipating new filings, and getting the filings, before transmitting the filing raw SGML to a Compute Instance. Detecting, anticipating, and getting, are all tasks that require little vcpu, so a worker pool should be able to do this on the cheap.

Fourth, we need to process the filing. I've written a fast secsgml parser in python, and a few months ago released a faster C implementation.

Besides parsing the SGML file into its constituent files (html, text, images), we also need to prepare it for the LLM. LLMs can take a wide variety of inputs, but in this case, we're going to restrict to text. Specifically, I am going to convert html, text, and pdf files directly to text format, then feed it into the LLM. Starting with a naive C implementation of doc2dict. Note that almost all SEC filings have text in them. Even when they appear to be all images, they have text, hidden inside, in a (fairly) nice usable form.

(4) will be handled by the Compute Instance. The instance will also be a websocket. So it contains user connection, and prompts and model user wants to use.

Fifth, after a filing is processed, we send it to OpenRouter with a model to use and prompts to use, e.g. Claude Sonnet 4.6 generate a three paragraph analysis.

This is likely to be input token cost heavy. For example, consider this 10-K Root Form. For the above prompt, the LLM would likely read ~100k tokens as input, and output about 600 tokens. At $3/mTok Input and $15/mTok output this would cost: $0.315. $0.30 for input, $0.015 for output.

This is where it gets fun. If we put the SEC filing text at the start of the message, then follow with a task such as 'generate a three paragraph analysis', we can use prompt caching, which is typically 90% cheaper for input tokens. This makes the first query cost $0.315, and future queries ~ $0.045.

So we now have a network effect. The more people use this product, the cheaper it becomes for everyone else.

Incidentally, this is (probably) a moat. Which investors like.

Sixth we transmit the result to user.

Maintenance

To make this as fast as possible, we probably need to have the websocket and processing code on the same compute instance. That means for updates, such as swapping in a better document parser, I may have to occasionally restart the instance, which would require users to reconnect.

Probably will do this on weekends when the SEC is not active. I also may need to setup an email chain to properly notify people.

Partners

The 8-K analyzer with one prompt should be fairly easy for me to setup, but bring your own prompt and model for all SEC filings is a bit tricky. Mostly, because processing that many tokens could be expensive if done wrong.

If you would like to be a part of a future pilot program, please email me at johnfriedman@datamule.xyz.