bilalcodes 👨🏻‍💻

I spent weeks building a free AI resume builder that actually beats ATS systems. Anyone can get an API to return text. The real engineering is everything that happens before and after. Here is the full architecture stack and the tradeoffs I made building س synapse. 🧵 (1/7)

@iannuttall @suraj1kc Tier 4 limits are brutal. I had to start aggressively routing tasks between Haiku and Sonnet depending on the complexity of the data extraction just to keep the limits in check. The architecture completely shifts once you hit that volume.

@suraj1kc using the claude api and even on tier 4 I'm hitting daily rate limits haha

Been quiet on here but I’m not dead. Although I’m sick so I might be soon? I’m rebuilding a niche site in Laravel with a focus on ecom and free tools! So far generated about 500k pieces of content using Claude (had to switch to Sonnet b/c Opus was $$$) The entire focus of this new site will be email subscribers from my free tools and then selling digital and possibly physical products via email later. I’ve got some incredible content now with charts and tables and… I’ve deleted a lot of old pages! This site wasn’t “hit” by HCU but I certainly branched out a bit too far into only loosely related topics. So I’m pulling it all back to the core topic that generates 90% of the traffic anyway (and adding another 100k pages that it didn’t have before) This is the same site I did a big migration on and lost pretty much all of the traffic for 6 months so once again it’s a YOLO event where it could go wrong. It’s making $2k/mo from ads but the plan is to replace that income with product sales and then I’ll remove all ads and maybe even go public with it. Final step is to rebuild sitemaps, test as much as I can with Screaming Frog and then launch it. Post-HCU this site gets 90k pageviews a month and I’ve deleted pages accounting for maybe 5-10k. I’m banking on the new, more relevant, pages more than making up for that! I know I’ve said this before, but this is the last pSEO migration I’m ever doing…

Handling this natively in Node.js is practically mandatory for production right now. I ended up building a custom round-robin rotation service for my own Claude orchestration just to keep the extraction pipeline alive during spikes. A single API key is a massive single point of failure at scale.

Another feature request for @conductor_build - allow me to load multiple Codex and Claude API keys and rotate through them when I hit the rate limits 🚀

Honestly, I didn’t set out to compete! I originally built it just to fix my own resume because standard HTML-to-PDF tools kept scrambling the text layer. I solved it using LaTeX under the hood, and then turned it into a full SaaS simply to learn the end-to-end product engineering lifecycle and understand market dynamics. Competing on strict machine readability is really just the byproduct of a learning project!

@bilalcodesdev This is good but how are you competing with so many similar tools in the market?

My feed is dead! Full of noise. No real people. No real conversations. I want to change that today. If you're out here building : → SaaS / AI tools → Vibe coding → Shipping in public → Figuring it out as you go Say Hi Drop what you're working on below.

@sans_builds Yep! Just pushed the V1 of the parsing engine live actually. It takes messy PDFs, extracts the data with Claude, and recompiles it into a perfectly ATS-readable format. You can test the live build here: resume.bilal.one

@bilalcodesdev Have you built anything with these tools yet? Just curious

@closermethod @elonmusk @AnthropicAI @SpaceX Rate limits are absolutely the final boss of shipping AI products right now. Moving from a local demo to a production environment where you have concurrent users hitting the API completely changes how you have to structure your backend.

The builders on Claude API feel this already. Rate limits are the current ceiling for what you can actually ship, not model quality. More compute infrastructure at scale means the agent workflows that only worked as demos last year become real products this year. Watching this closely.

As the recently expanded partnership with @AnthropicAI demonstrates, @SpaceX is offering AI compute as a service at significant scale. We are in discussions with other companies to do the same. Over time, especially with orbital data centers, we expect to serve AI at extremely high scale.

It is the classic gym membership model! The Pro subscription relies on the fact that 90% of users won't max out their limits, which subsidizes the 10% who do. With the API, you are paying for guaranteed, raw compute on demand. To keep the burn rate down on my own SaaS, I had to aggressively route tasks between Haiku and Sonnet depending on the complexity of the data extraction.

Genuine question: $20 of Claude API can vanish in minutes with heavy usage, but $20 for Claude Pro lasts much longer because of rate limits + rolling windows. So what’s the actual economics here? Is Claude Pro effectively giving users far more value than $20 worth of API credits? If yes, how do startups building on Claude’s API make sustainable margins when users could just pay Anthropic directly for Pro instead?

Exactly this. When I was architecting the backend for Synapse, I quickly realized you can't just pipe the Claude API directly to the client. I had to build a dedicated queueing and retry layer in Node just to handle rate limit spikes and auth blips gracefully. If an agent doesn't have a robust fallback mechanism, it’s just a demo.

First job: insurance. Claude API has 98.99% uptime over 90 days. Sounds solid until you do the math: that is 21 hours of potential downtime. Rate limits, auth failures, cost spikes, regional issues. Your agent does not see the difference. It just stops.

My AI agent failed at 4:17 AM. Primary API was down. The fallback triggered. Cost me $0.02 that month. That $20 in OpenRouter credits does two jobs. I only planned for one.

@justinhammon_ Appreciate the shoutout, Justin! Always great to cross paths with another builder tackling the ATS puzzle.

@bilalcodesdev This is super helpful!

ATS doesn't read your resume. It parses it. If your formatting confuses the parser - your excellent career gets a bad score before a human ever looks. Use: → Simple formatting → Exact keywords from the JD → Clean PDFs (no tables, text boxes, graphics) This alone can fix a silent rejection problem.

@Umesh__digital @Francklp2 For sure 🙌🏻

@bilalcodesdev @Francklp2 lets connect..

Hey founders, looking to #connect with people in: SaaS AI tools AI agents Growth & distribution Web apps Tools that solve your own problem Anything that sounds too niche to work ⁉️ Drop your live project or what you're working on 👇

@BenjDicken Tell me about it 😂. Running complex ATS parsing through Claude for my engine completely flipped my cloud budget. My AWS infrastructure bill is basically a rounding error compared to my Anthropic API usage now.

Plug in your anthropic key and it becomes opex

@aminnnn_09 @Priyannkaaaa Spot on. Complex layouts almost always scramble the text layer for the parser. If they need a format built specifically to pass the extraction phase cleanly, I built a free engine that compiles parser-safe PDFs: resume.bilal.one

@Priyannkaaaa Nope ATS parser will reject this

Would this resume get shortlisted if Sam Altman applied for a Co-Founder role at Anthropic? 👀

This is the exact headache I ran into when building my parser. Workday and Greenhouse process text layers completely differently when they encounter multi-column PDFs. I ended up bypassing HTML-to-PDF entirely and using a custom Node.js service to compile LaTeX. It forces a linear text layer that both Workday and Greenhouse can read flawlessly, regardless of the visual layout.

@faizan_raza1350 @expo mobile-first ATS is sharp — most resume tools default to desktop. curious how you handle the parser quirks: have you tested against Workday vs Greenhouse output? those two diverge enough that 'ATS-friendly' is really two different formats.

🙋‍♂️We'd like to connect with developers who have built production grade mobile apps with Expo + Claude Code OR Codex. Can you give us a shout if this is you?

Great prompt! I actually got so tired of doing this manually that I orchestrated this exact workflow into a free tool. It uses Claude Haiku to extract the data and scores it directly against the pasted JD to highlight the exact keyword gaps. If any of your followers want the automated version, they can run it here: resume.bilal.one

6. Score it before you send it. Claude prompt: "Act as an ATS parser. Score my resume against this job description out of 100. Break the score into keyword match, formatting readability, and experience alignment. Tell me the three changes that would move the score the most." Under 80, fix it. Under 90, fix it more. Most candidates send at 50 to 60 and wonder why nothing comes back. This whole process took me 3 hours. Without Claude it would have taken 3 weeks of trial and error and a career coach charging $200 an hour. Keep one master resume with every bullet, project, and metric you have ever produced. For each new job, copy it, ask Claude to tailor it, submit within an hour. The candidate who applies first with a 90% match beats the candidate who applies third with a 70%.

Last week, I spent 3 hours with Claude rebuilding my resume. Result: 14 missing keywords added. 6 weak bullets rewritten. 92% ATS match score on the next application. How I did it, step by step. Save this:

Canva is great for design, but terrible for data extraction. The text layers just turn to garbage. If any of your followers are stuck in this loop, I built a free tool that pulls their data and compiles it into a strictly ATS-compliant format so they stop getting auto-rejected: resume.bilal.one

Most candidates are still using Canva to design their CVs, ATS cannot parse Canva. Your CV is being rejected before a human even sees it.

Spot on, Phillip. Multi-column PDFs from design tools completely scramble the underlying text layer. I actually got so frustrated by this that I built an engine that compiles resumes using LaTeX. It keeps the visual layout clean for humans but enforces a perfect, linear text layer for ATS scanners. Free to use if you ever need to point people toward a parser-safe option: resume.bilal.one

Multi-column layouts, tables, and skill bars break ATS parsing on nearly every major platform. The creative CV that impresses at a career fair is the same format that fails silently when the application goes in. Two audiences. Same document. Opposite outcome.

@Tame_abeast Huge milestone, congratulations on unlocking the monetization!

5M impressions secured 🙌📊 The feelings when you finally monetised 😌😌😌🎉🎉🎉🎉 I accept congratulations 🎊 🎊🎊🥳🥳🥳🥳

@vanshyadav1408 This is exactly why I kept س synapse closed-source but free to use. The moment you open the repo for a fully functional SaaS tool, you aren't building a community—you are just doing the heavy lifting for someone else's Stripe integration.

Open source is fun until someone fork and add a paywall.

@olivercingl The 'drop your link' economy is booming 😂. It’s getting harder to find actual engineering discussions buried under all the engagement farming.

what is happening to my feed