ok been quietly working on this thing for the past 5 months, and the reports finally hit the quality bar i was aiming for.
long story short: i built a system that scrapes a company’s site, profiles, press, and anything else it can find online, then runs it through a multi-model prompt engine i wrote from scratch (multiple LLMs chained together with custom logic). the workflow is split into three parts: crawl, process, synthesize.
output comes in markdown. all signal, no bs. useful for GTM, growth, competitor analysis, product gaps, strategic blindspots, that kind of thing.
i’ve been stress testing it on real companies. just pushed 2 sample reports to GitHub if you wanna peek:
not selling it as a product (yet). backend still in dev. just doing reports manually for now while testing edge cases/validating if anyone wants this at all lol.
note: it’s not fully automated. i use the system to generate raw analysis, then apply human input to reframe, expand, or clean up weak insights. think of it like pairing with an AI researcher who works fast but lacks nuance/context, you still need a strategist to shape the output. there's tons more feature but at the core this is it.
happy to run one or two more this week if anyone wants a teardown. not free though (token cost’s brutal lol). DM if interested.
Agreed. A lot of SEA countries been handling it pretty well(with the exception of Indonesia). I'm from Malaysia and even though there's been a huge political storm recently, we still have the highest recovery rate among all fo the countries that are suffering.
This is funny, if you actually google the keyword X-Keyscore, you'll find job opportunities that matches the criteria of intelligence gathering. And if you look closely, you'll find out it's a company called Raytheon that's awarded the contract to execute these works. And they have main offices in these crucial locations, eg; Fort Meade, Australia etc. And that they do all these kind of intelligence works.
And ironically, Raytheon's scientist was the dude who invented microwave.
What happens next looks like a 21st-century data assembly line. At the NSA’s headquarters in Fort Meade, Maryland, a program called Xkeyscore processes all intercepted electronic signals before sending them to different “production lines” that deal with specific issues. Here, we find another array of code names.
Pinwale is the main NSA database for recorded signals intercepts, the authors report. Within it, there are various keyword compartments, which the NSA calls “selectors.” Metadata (things like the “To” and “From” field on an e-mail) is stored in a database called Marina. It generally stays there for five years. In a database called Maui there is “finished reporting,” the transcripts and analysis of calls. (Metadata never goes here, the authors found.)
As all this is happening, there are dozens of other NSA signals activity lines, called SIGADS, processing data. There’s Anchory, an all-source database for communications intelligence; Homebase, which lets NSA analysts coordinate their searches based on priorities set by the Director of National Intelligence; Airgap, which deals with missions that are a priority for the Department of Defense; Wrangler, an electronic intelligence line; Tinman, which handles air warning and surveillance; and more.
Lest you get confused by this swirl of code names and acronyms, keep this image in mind of the NSA as a data-analysis factory. Based on my own reporting, the agency is collecting so much information every day that without a regimented, factory-like system, analysts would never have the chance to look at it all. Indeed, they don’t analyze much of it. Computers handle a chunk, but a lot of information remains stored for future analysis.
long story short: i built a system that scrapes a company’s site, profiles, press, and anything else it can find online, then runs it through a multi-model prompt engine i wrote from scratch (multiple LLMs chained together with custom logic). the workflow is split into three parts: crawl, process, synthesize.
output comes in markdown. all signal, no bs. useful for GTM, growth, competitor analysis, product gaps, strategic blindspots, that kind of thing.
i’ve been stress testing it on real companies. just pushed 2 sample reports to GitHub if you wanna peek:
- Grab (SEA unicorn) - LemmyHomes (early-stage Malaysian proptech)
→ https://github.com/lightyoruichi/numen-output
not selling it as a product (yet). backend still in dev. just doing reports manually for now while testing edge cases/validating if anyone wants this at all lol.
note: it’s not fully automated. i use the system to generate raw analysis, then apply human input to reframe, expand, or clean up weak insights. think of it like pairing with an AI researcher who works fast but lacks nuance/context, you still need a strategist to shape the output. there's tons more feature but at the core this is it.
happy to run one or two more this week if anyone wants a teardown. not free though (token cost’s brutal lol). DM if interested.