Project summary

I got tired of watching my LangChain scripts crash and get duped when wandering around the internet. Basically, when you let an autonomous agent fetch web pages, you run headfirst into two annoying headaches.

Either the bot stumbles into a malicious crypto drainer or brand-new phishing link that was registered 40 minutes ago (which Google Safe Browsing completely misses because it takes days to index), or it visits a completely legitimate site like Stripe or Cloudflare and crashes because someone updated a CSS class name and BeautifulSoup threw a NoneType error.

So I wired up two small engines to fix both sides of this. The first engine is PhishVision. Before any agent opens an unknown URL, PhishVision does an aggressive heuristic and visual scan in about a second and a half to see if it's spoofing a login or wallet drainer. In my own tests running 50 fresh zero-day scam links from OpenPhish, it caught 34 right off the bat while Safe Browsing caught zero. The second engine is OpticParse. Instead of parsing messy DOM trees that change every week, it takes a screenshot and uses vision to extract the fields you actually asked for. It handles dynamic React hydration and bot walls without breaking.

What are this project's goals? How will you achieve them?

Right now my catch rate on live 0-day scam links sits at 68%. I want to push that past 85%.

The links that slip past are almost always compromised WordPress blogs with clean 8-year-old domain reputations where an attacker shoved a credential stealer inside an obscure subfolder. To fix that without turning the scan into a sluggish 20-second crawl, I'm building out a lightweight HTML DOM inspector that flags hidden credential forms and wallet connection scripts on the fly.

Beyond that, my main goal is keeping the integrations completely free and accessible for developers building on LangChain and LlamaIndex. I also want to keep publishing our flagged scam datasets directly on Hugging Face so independent safety researchers can study real adversarial agent attacks instead of synthetic benchmarks.

How will this funding be used?

I need 3500 dollars to pay for server bills for 6 months.

about 1200 dollars goes to rotating residential proxies because when our scanner checks shady scam sites they ip ban us quickly if we use the same ip.

about 1300 dollars is for render and cloudflare servers so we can run headless chromium browser checks without running out of ram.

and 1000 dollars is for cloudflare r2 storage and data bandwidth to host the free threat datasets on huggingface for the community.

Who is on your team? What's your track record on similar projects?

I am a solo developer based in India building all of this independently. I don't come from a big tech lab or a fancy university; I taught myself how to build scalable systems by hacking on projects and paying for servers out of pocket.

Over the past few months I have written and published the opticparse-py, langchain-opticparse, and llama-index-tools-opticparse libraries on PyPI, created MCP tools for Claude Desktop, released open phishing datasets on Hugging Face under paras9909, and architected an edge worker network that has handled hundreds of thousands of requests with zero downtime. All the client code and benchmark reproductions are open on my GitHub at https://github.com/parastejpal987-cmyk/opticparse-public.

What are the most likely causes and outcomes if this project fails?

The biggest engineering hurdle is balancing detection depth with API latency. If an analysis takes 25 seconds because of heavy JavaScript execution, developers will just disable the security check to make their bots run faster. If we cut it down to 500 milliseconds, sophisticated evasion kits might slip through. Finding that exact sweet spot is the hardest part.

Even if developer adoption is slower than anticipated, the datasets, benchmark suites, and open-source packages are already out in the wild and will remain completely accessible for the AI safety community.

How much money have you raised in the last 12 months, and from where?

0 USD. Everything so far has been self-funded using my personal savings, free-tier hosting limits, and a twelve-dollar domain name.