Alexander Zhang

I build an autonomous research fleet that reads the AI-security literature so a person doesn’t have to.

What it is

A continuously running multi-agent system. Several unattended jobs sweep research forums, preprint servers, issue trackers, and opportunity boards on a fixed schedule, judge what they find against a maintained corpus, and surface only what is genuinely new.

It has been running since July 2026. It maintains a corpus of roughly 900 catalogued findings and an archive of generated research directions, each one checked against the published literature before it is reported. Most are rejected — that is the point.

How it works

• Sweep. Scheduled jobs read a registry of vetted sources across several independent fetch channels, honouring each site's robots.txt and terms.
• Ground. Findings are recorded with a verbatim quote, a date, and a source URL, so an entry survives its source going offline.
• Generate. A separate job proposes research directions from the corpus through a set of fixed analytical lenses.
• Kill. Every proposal is searched against the published literature. Anything already solved is archived with the citation that killed it.
• Report. What survives is written up with what is measured and what is still unverified.

Principles

• Terms and protocols are followed, never worked around. No user-agent spoofing, no CAPTCHA solving, no proxy laundering. A robots.txt disallow is final.
• Nothing time-varying is asserted from memory. Prices, deadlines, and access rules are re-checked at the source and cited with a date.
• Every claim carries its evidence, and unverified claims are labelled as such rather than rounded to something plausible.
• Depth over volume. A single well-grounded finding beats a page of summaries.

Status

Independent research project, self-funded, run by one person. Not a company, no product for sale, no customers, no revenue. Built to answer research questions and to find opportunities worth a person's time.

---