youchao

[youchao @iceworldchamplo]

Live Demo: https://t.co/1vK0xOWeMl

Source code: https://t.co/ZOQsAKlQbG

[youchao @iceworldchamplo]

Data ingestion from PDFs, images, complex documents and HTML has always been a hard step.
Accurately recognizing handwriting, LaTeX formulas and tables directly affects training quality, RAG knowledge bases and ontology stacks.

[youchao @iceworldchamplo]

At enterprise scale, NVIDIA cuVS can build HNSW graphs on the GPU, speeding up ingest and reindexing of large vector datasets (check version support):
https://t.co/1z2eyoKzYp

Would love to hear which models you'd add to the comparison.

#OCR #RAG #DocumentAI #LLM #Elasticsearch

[youchao @iceworldchamplo]

• Agentic search over the knowledge base: OCR chunks + embeddings + metadata indexed in Elasticsearch, using hybrid retrieval (BM25 + kNN merged with Reciprocal Rank Fusion) rather than pure vector search.

[youchao @iceworldchamplo]

What I'm exploring next
• OCR for web pages
• Other inference serving options: RunPod (containers and serverless), or self-hosting with llm-d + KEDA + vLLM

[youchao @iceworldchamplo]

Chandra OCR 2 produced the cleanest table layout, but clean layout is not the same as a complete reading: on page 5 it draws the grid and leaves nearly every cell empty — 21,033 characters, of which about 262 are text.

[Marry Evan @marryevan999]

I was woken up at 3:47 AM by OpenClaw

It sent just one message:
"Found 6 markets that will settle in the next 90 minutes. Americans are still asleep. Need approval to deploy $12K."
I replied with a yes and went back to sleep

Woke up in the morning, and my account had gained:
+$43,800
I've been running this agent for 9 days
It does one thing specifically:
Watch for timezone arbitrage

I fed OpenClaw a few types of real-time feeds from different time zones:
Japan government RSS
European Parliament schedule
Australian financial alerts
Middle East flight tracking
Asian central bank announcements

Then I gave it just one rule:
"Find markets that settle between 2 AM and 6 AM Eastern Time. If the edge exceeds 30%, wake me up."
And at 3:47 AM, it actually found 6 markets
All settling between 4 AM - 6 AM

These markets had one thing in common:
The market was still pricing on a "normal rhythm"
But when settlement happened, US traders were basically all asleep
The official signals from the relevant countries had actually come out early

The alerts it pushed to me at the time were:
"Japan rate decision - BOJ leak shows YES 68%, Polymarket still at 23¢"

"EU emergency vote - Live footage shows YES already leading, Polymarket still at 31¢"

"South Korea policy - Government RSS has confirmed, Polymarket still at 19¢"
"Australia trade deal - Minister stated publicly 2 hours ago, Polymarket still at 27¢"
"UAE production cut - OPEC meeting minutes already public, Polymarket still at 15¢"

"Singapore regulation - Parliament session still live-streaming, Polymarket still at 22¢"

Its summary was pretty straightforward:
Potential edge: $43K
Window: 90 minutes
Required capital: $12,000
I was half-asleep at the time, phone buzzed once

Opened Telegram and saw just one line:
"approve or miss"
I replied yes, then went back to sleep
By 7:30 AM when I woke up, all the notifications had come in

All 6 markets settled during morning hours in Asia / Europe
While US traders were waking up, the markets were already done

My entry prices were roughly:
15¢ - 31¢

Final settlements all hit:
95¢ - 100¢

Profit breakdown:
Japan: $8,200
EU: $6,900
Korea: $11,400
Australia: $7,100
UAE: $5,800
Singapore: $4,400

Total:
+$43,800

Later when I checked the logs, I realized this agent had been monitoring these markets for 8 to 14 hours
Constantly syncing official sources
Constantly waiting for US traders to go to sleep

Then it only struck in that instant:
Results overseas were basically confirmed
Prices on the US side hadn't updated yet
And settlement was already close

This edge boils down to something pretty simple:
Polymarket is 70% US traders
But events around the world never happen on EST time

While you're sleeping, the markets keep settling
This play of specifically exploiting info gaps during "when Americans are asleep" hours—do you think it's timezone arbitrage, or is it edging into the most basic form of insider advantage?

Giving This Free for 24 hours. To get it:

1. Comment the word 'Openclaw'
2. Like and Retweet this post
3. Follow me @marryevan999 (so i can DM you)

[youchao @iceworldchamplo]

@marryevan999 Openclaw

[youchao @iceworldchamplo]

@AlexFinn hermes carry llm-wiki skill that build knowledge base for obsidian to ingest easily

[youchao @iceworldchamplo]

@AlexFinn I ran hermes-agent in hostinger VPS docker containers and installed codex or claude code in VPS host machine as a backup repair fixer. I heard the docker limits its ability especially maybe in browser visual clicking simulation, daily back up still needed for safety.

[youchao @iceworldchamplo]

@ErickSky isn't this my roborock vacuum cleaner doing many years ago, what is new about this

[youchao @iceworldchamplo]

@NicoPrincely @frankie40202099 @fuhknjo not everyone has the intelligence to make bombs or poison gas without being noticed by neighbors. In this case, do you think a high school student can easily become Walter White? They may kill themselves when experimenting. Using knives do you think they can kill 20?

[youchao @iceworldchamplo]

@NicoPrincely @frankie40202099 @fuhknjo Evil people are in every country, in many countries they can only get knives, in others they can only get a pistol. They were taken out by police and never saw such great loss happens. We should stop lying to ourselves that these heavy weapons are not the cause.