Soo

AI Researcher, Healthcare, Computer Vision, LLM

[Soo @sooperset]

@akshay_pachaar This is also good mcp: https://t.co/8NG43JEIbp

[Soo @sooperset]

@adamsilverman @AgentOpsAI Can’t believe you missed https://t.co/8NG43JEalR

[Soo @sooperset]

@PawelHuryn Thanks for using mcp-atlassian! I'm happy to see this demo as the author. Anyone can use this: https://t.co/Y4uoJ2te1N

[Soo @sooperset]

Built an MCP that connects LLM to Jira and Confluence. Now LLM can create tickets, search docs, and update pages.

With 168K downloads, it's the 3rd most popular community MCP. Check out this demo where it creates Jira tickets from meeting notes!

https://t.co/8NG43JEIbp

[Soo @sooperset]

@Saboo_Shubham_ Confluence and Jira also accessible! https://t.co/t30R35GqQ9

[Soo @sooperset]

Built an MCP that connects LLM to Jira and Confluence. Now LLM can create tickets, search docs, and update pages.

With 168K downloads, it's the 3rd most popular community MCP. Check out this demo where it creates Jira tickets from meeting notes!

https://t.co/8NG43JEIbp

[Soo @sooperset]

@svpino MCP should not be the openapi wrapper. I’ve made Jira and confluence MCP and realized that MCP sever is more than OpenAPI wrapper: https://t.co/t30R35FT0B

[Soo @sooperset]

Built an MCP that connects LLM to Jira and Confluence. Now LLM can create tickets, search docs, and update pages.

With 168K downloads, it's the 3rd most popular community MCP. Check out this demo where it creates Jira tickets from meeting notes!

https://t.co/8NG43JEIbp

[Soo @sooperset]

@sundarpichai Hope one day I can manage Jira and confluence projects via google studio: https://t.co/t30R35GqQ9

[Soo @sooperset]

Built an MCP that connects LLM to Jira and Confluence. Now LLM can create tickets, search docs, and update pages.

With 168K downloads, it's the 3rd most popular community MCP. Check out this demo where it creates Jira tickets from meeting notes!

https://t.co/8NG43JEIbp

[Cloudflare @Cloudflare]

You can now build and deploy remote MCP servers to Cloudflare, and we handle the hard parts of building remote MCP servers for you. https://t.co/M4vZfDHvS9

[Soo @sooperset]

@OpenAIDevs now we can integrate jira and confluence with chatgpt!

https://t.co/8NG43JEIbp

[Soo @sooperset]

@sama Now your jira can be managed by chatgpt!

https://t.co/Y4uoJ2te1N

[Soo @sooperset]

Built an MCP that connects LLM to Jira and Confluence. Now LLM can create tickets, search docs, and update pages.

With 168K downloads, it's the 3rd most popular community MCP. Check out this demo where it creates Jira tickets from meeting notes!

https://t.co/8NG43JEIbp

[Soo @sooperset]

@_avichawla I'd like to introduce mcp-atlassian which connects Jira and Confluence with AI models. With 168K downloads (3rd most popular community mcp), it lets LLM access your confluence and jira content.

https://t.co/8NG43JEIbp

[Soo @sooperset]

Built an MCP that connects LLM to Jira and Confluence. Now LLM can create tickets, search docs, and update pages.

With 168K downloads, it's the 3rd most popular community MCP. Check out this demo where it creates Jira tickets from meeting notes!

https://t.co/8NG43JEIbp

[Woojin Kim @woojinrad]

‼️Why clinical domain expertise in #radiologyAI research is important! Even without adversarial conditions, general-purpose #LMMs are unreliable for medical diagnosis. While I am unsurprised by the results, you also CANNOT evaluate with faulty data sets. https://t.co/4Mu9AuT6gS

[Soo @sooperset]

https://t.co/O2Szsi9gW6

[Soo @sooperset]

𝐕𝐢𝐓𝐀𝐑: 𝐕𝐢𝐬𝐢𝐨𝐧 𝐓𝐫𝐚𝐧𝐬𝐟𝐨𝐫𝐦𝐞𝐫 𝐰𝐢𝐭𝐡 𝐀𝐧𝐲 𝐑𝐞𝐬𝐨𝐥𝐮𝐭𝐢𝐨𝐧

How can Vision Transformers (ViTs) be adapted to handle variable input resolutions while maintaining high performance and computational efficiency?

The authors propose Vision Transformer with Any Resolution (ViTAR), which introduces an Adaptive Token Merger (ATM) module for dynamic resolution adjustment and Fuzzy Positional Encoding (FPE) to provide consistent positional awareness across multiple resolutions.

[Soo @sooperset]

https://t.co/yw04cqWi7q

[Soo @sooperset]

ViTamin: Designing Scalable Vision Models in the Vision-Language Era

𝑆𝑜𝑚𝑒 𝑘𝑒𝑦 𝑖𝑛𝑠𝑖𝑔ℎ𝑡𝑠:
• 𝐃𝐚𝐭𝐚 𝐒𝐜𝐚𝐥𝐚𝐛𝐢𝐥𝐢𝐭𝐲: As training data increases, performance consistently improves across all models
• 𝐌𝐨𝐝𝐞𝐥 𝐒𝐜𝐚𝐥𝐚𝐛𝐢𝐥𝐢𝐭𝐲: ViT demonstrates the most effective scaling with model size
• 𝐅𝐞𝐚𝐭𝐮𝐫𝐞 𝐑𝐞𝐬𝐨𝐥𝐮𝐭𝐢𝐨𝐧: The final feature map resolution affects prediction accuracy
• 𝐇𝐲𝐛𝐫𝐢𝐝 𝐀𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐞: Hybrid models like CoAtNet generally outperform pure CNN or transformers

[Soo @sooperset]

Retrieval-Augmented Dual Instruction Tuning (RA-DIT) is a powerful new approach for building RALMs through lightweight fine-tuning.

By updating the LM and retriever separately on carefully selected datasets, RA-DIT significantly boosts zero- and few-shot performance on knowledge-intensive tasks. RA-DIT 65B sets a new bar, outperforming existing open RALMs and approaching the capabilities of larger closed models.

https://t.co/O1uEYKKFSs

[Soo @sooperset]

Move over GPT-3.5 - @databricks' new open LLM DBRX is here, and it's breaking records.

https://t.co/QwnAVvczvE

[Soo @sooperset]

Some key takeaways and future directions:
• The developed pipeline successfully integrates AI findings directly into radiologists' structured reporting workflow in an automated way
• This enables time savings for radiologists compared to manually adding AI results to free-text or structured reports
• The approach facilitates storing validated AI findings in structured databases for later analysis and research
Extending the pipeline to more AI tools and examination types is an important next step
• Increasing adoption of structured reporting standards will further support the clinical utility of this AI integration method

[Soo @sooperset]

Title: A novel reporting workflow for automated integration of artificial intelligence results into structured radiology reports

https://t.co/H4HonBpqyU

Key points:
• Pipeline automatically populates structured report templates with AI findings
• Aims to provide a more standardized, efficient way to leverage AI
• Evaluated using a commercial AI tool for chest X-ray pathology detection

The researchers assessed the pipeline's performance by comparing reporting times and quality to free-text reports and conventional structured reports. The approach shows promise for streamlining clinical AI use.

[Soo @sooperset]

In addition to speed, the study assessed report quality on a 5-point scale for completeness and correctness.

Pipeline-generated structured reports scored significantly higher than free-text reports (p<0.001) (Fig 6)

There were no significant differences between the pipeline reports and conventionally created structured reports. (Table 2)

This suggests the AI-to-Structured Reporting approach can maintain report quality while increasing efficiency.