Update: I made it to the final 20. ๐บ๐ณ
UN ESCAP's global hackathon on AI for regulatory analysis.
Team size: still 1.
Here's what I built, and the one decision I almost didn't make. ๐งต
1/
Format: open round โ 40 teams โ final 20 โ live pitch on 3 Aug.
I entered solo. Then I got sick. Total working time: 8 days.
2/
ReguLink crawls official government legal portals, reads thousands of pages of statute, and outputs every relevant provision โ with a verbatim quote and source URL for each.
3 economies. 290 rows of provision-level evidence.
Total AI cost: $3.
3/
Low cost comes from where the money doesn't go:
โ Rule pre-filter drops ~70% of the text. Free.
โ Self-hosted MiniLM ranks the rest. 90MB, CPU only. Free.
โ Only the top 12 chunks reach a paid model.
Most RAG spend is reading things you didn't need to read.
4/
The real risk with LLMs reading law isn't accuracy. It's fabricated citations.
A quote that looks like statute but isn't.
So every quote is searched back into the source, character by character. Three tiers: exact โ normalised โ fuzzy at 0.92.
Fail all three? Rejected. Not scored.
5/
It caught a real one.
The model quoted a provision but silently dropped four words. Reads perfectly. Changes the legal meaning.
Match score: 0.798. Below threshold. Rejected.
That's the whole system in one number.
6/
The decision I almost didn't make:
I hand-audited 63 provisions and found a high false-positive rate on one indicator.
I could have quietly left it out.
I tagged every one and shipped it in the submission.
7/
A compliance tool that hides its own failures isn't a compliance tool.
Same reason I documented a retrieval gap that didn't even change my final score โ the score was right, but the step was broken.
A correct score can still hide a broken step.
8/
Live pitch: 3 Aug. 8 min talk, 7 min judge Q&A, in English.
English isn't my first language and my listening is worse than my reading.
Currently drilling a script. ๐
Results 5 Aug.
9/
Independent developer. I take on AI application work, automation pipelines, and data engineering.
DMs open.
#buildinpublic #AI #RAG #LegalTech #RegTech #indiedev
Iโve been locked out of my Claude account and my export link is expired/invalid. All my project files for the UN ESCAP Global Hackathon 2026 (where my team is shortlisted Top 40) are trapped inside. Final submission deadline is July 20 โ less than 3 weeks away. I already submitted an appeal + export request, but got no resolution. This is extremely irresponsible for such a time-critical international competition. Please help restore access or send a working export immediately. Hundreds of hours of work at risk.
Account email:
[email protected]
I just tried logging into the last remaining account โ it also failed completely.
Now I have **zero access** to any of my accounts.
No warning. No working export. No way to retrieve my UN Hackathon project.
This is beyond frustrating. My team is shortlisted for the UN ESCAP Global Hackathon (due July 20) and all our work is gone because of your sudden bans.
@AnthropicAI@claudeai Please respond and fix this ASAP.
#Claude #Anthropic
@claudeai@AnthropicAI
I just tried logging into the last remaining account โ it also failed completely.
Now I have **zero access** to any of my accounts.
No warning. No working export. No way to retrieve my UN Hackathon project.
This is beyond frustrating. My team is shortlisted for the UN ESCAP Global Hackathon (due July 20) and all our work is gone because of your sudden bans.
@AnthropicAI@claudeai Please respond and fix this ASAP.
#Claude #Anthropic#
@AnthropicAI
This is outrageous:
- 3 accounts suddenly suspended with **zero prior warning**.
- One account can't even be logged into at all.
- For the others, the data export links you sent expired immediately and now show "Invalid link".
I have **zero access** to my projects for the UN ESCAP Global Hackathon 2026 (shortlisted Top 40, final submission due July 20).
Months of work on an official United Nations competition โ completely trapped because of your sudden bans and broken export system.
Fix this immediately.
#Claude #Anthropic
@AnthropicAI
Iโve been locked out of my Claude account and my export link is expired/invalid.
All my project files for the UN ESCAP Global Hackathon 2026 (where my team is shortlisted Top 40) are trapped inside. Final submission deadline is July 20 โ less than 3 weeks away.
I already submitted an appeal + export request, but got no resolution. This is extremely irresponsible for such a time-critical international competition.
Please help restore access or send a working export immediately. Hundreds of hours of work at risk.
#Claude #Anthropic #UNHackathon
There was zero warning or prior notification before all three of my accounts were suddenly suspended.
I had no chance to export my data beforehand. When I requested an export, the โYour data is ready for downloadโ links you sent are already expired and invalid.
This sudden ban with no export window is extremely unfair and irresponsible.
Please restore access or provide working export links immediately.
Haven't posted on X for half a month, time for a quick update:
Been battling a hyperthyroidism relapse, mostly lying in bed recovering. Then Claude decided to join the party and banned all three of my accounts in one go...
Apparently even people who never used it are getting ban notices now. The "you used me or not, you're banned" policy ๐
Life is wild sometimes โ illness hits like a truck, bans come like a silk thread...
Taking it slow with recovery and projects. Anyone else dealing with this? Let's commiserate.
#ClaudeBanWave #HyperthyroidismLife #DeveloperStruggles
On May 31, I received an email that meant a lot to me.
My project was selected for the shortlist of the UN Global Hackathon on Using AI for Digital Trade Regulatory Analysis, among 129 teams from 30 countries.
As a self-taught indie developer without a computer science degree or big-tech background, seeing that email was both exciting and surreal.
But the real challenge started after the acceptance.
Over the past few days, I've been studying the competition framework, regulatory datasets, and evaluation criteria while building an AI research agent that can crawl, parse, and analyze digital trade regulations across different jurisdictions.
Today, after an entire day of debugging, I finally got the Singapore legislation pipeline running end-to-end.
Along the way I discovered that:
โข The main legislation page doesn't contain the actual provision text
โข The real content is hidden behind a specific provision endpoint
โข Sections and subsections require custom parsing
โข Multiple date fields can break metadata extraction
โข A 200 OK response doesn't always mean you have valid content
โข Status pages can easily masquerade as real legislation pages
The funny part is that a few months ago, I probably would have assumed:
โข My VPS was blocked
โข My IP was banned
โข I needed residential proxies
โข I needed Playwright and stealth tooling
โข I needed another dozen scraping libraries
Instead, the solution turned out to be much simpler:
A correct URL is often worth more than ten proxies.
One lesson I've learned as a developer is that the biggest obstacle is rarely technical complexity itself โ it's the wrong assumptions we make before investigating the problem.
Back to building. More countries to support, more regulations to analyze, and a lot more work ahead.
What a week โ
Joined a UN hackathon, got a "job interview" in the middle of it, realized it was a scam when they asked me to run a ZIP file locally. Closed it and went back to coding ๐
Just submitted ReguLink Asia โ an AI-powered Asia-Pacific regulatory compliance platform. Solo dev, built from scratch.
Form filled. Video compressed (3 times ๐ฉ). Uploaded. Done.
Whatever happens, I shipped it. โ
#DevNetwork #Hackathon2026 #BuildInPublic #IndieHacker
Built a production-grade LLM resilience system in 1 day ๐ฅ
SkinSense AI โ AI skin analysis that NEVER goes down.
3-layer fallback architecture:
โ Layer 1: DeepSeek (primary)
โ Layer 2: Kimi (auto-failover)
โ Layer 3: Rule-Based Engine (always on, offline)
Even when ALL AI providers fail simultaneously, users still get personalized skincare recommendations based on their actual skin data.
Real-time Dashboard shows routing chain, latency & cost savings vs GPT-4.
๐ Demo: https://t.co/tjAepV1grQ
๐ Dashboard: https://t.co/f6sP8yT8yp
๐ป GitHub: https://t.co/5uGkJd6KES
Built for @DevNetwork AI+ML Hackathon 2026
#buildinpublic #indiehacker #LLM #AI #NextJS
@CricTalk29 Hey! Built LicenseKit โ license key management for indie developers. โ Auto-delivery after payment โ Device binding (prevent sharing) โ Real-time verify API โ Multi-project support Free to start, no credit card needed. https://t.co/lRKxlncfhC Would love your feedback! ๐
The hardest part of building solo isn't coding.
It's UI design.
Every morning I open my laptop and think:
"Does this look like a $129 product or a $9 template?"
No design background.
No budget for a designer.
Just me, Tailwind, and way too many hours
staring at padding values.
Shipping anyway.
#buildinpublic #indiehacker #solofounder
What I've been building lately:
1/ ArmorStack โ Next.js boilerplate with
Stripe + WeChat + Alipay. Shipped in 4 hours.
โ https://t.co/eKxMBhnKdf
2/ IELTS Quanku โ Full IELTS prep platform.
Vocab, listening, writing, speaking, all in one.
โ https://t.co/XWSfQRNvnF
3/ English Pro โ Chinese English learners platform.
Phonics to IELTS, complete learning path.
All built solo. All shipped in the last 2.5 months.
Started from zero code knowledge.
Just keep shipping.
#buildinpublic #indiehacker
@ilan_hertz The 'Stop fighting CSS box-shadow' part hit me hard. I think weโve all spent way too much time in DevTools just for a shadow. ๐ Great tool!