Sarthak Chauhan

Building | Prev: IIT Delhi

[FreightWaves @FreightWaves]

Lanesurf is redefining carrier outreach. Its AI books a load in under 10 minutes by speaking to dozens of carriers at once across calls, email and text. It vets for compliance, negotiates rates in real time and delivers the best carrier at the best price—turning hours of manual work into minutes. Freight brokers are using Lanesurf to boost speed, accuracy and margin at scale. https://t.co/I3L58ekLUJ

[Michael Truell @mntruell]

A conversation with @patrickc on old programming languages, software at industrial scale, and AI's effect on economics/biology/Patrick's daily life.

00:15 - Why Patrick wrote his first startup in Smalltalk
03:35 - LISP chatbots
06:09 - Good ideas from esoteric programming languages
09:12 - Brett Victor and Dynamicland
16:37 - Programming human organizations
20:28 - A codebase's "Big Bang" moment and MongoDB
25:49 - Rewriting Stripe
32:00 - How do you, Patrick Collison, use AI?
38:25 - Changes to GDP/TFP
41:56 - Programming human biology
46:10 - Unexpected beneficiaries of AI

[Chroma @trychroma]

Introducing our latest technical report: Context Rot - How Increasing Input Tokens Impacts LLM Performance

Our results reveal that models do not use their context uniformly.

full report in replies

[kelly @kellyhongsn]

excited to share my latest technical report @trychroma!

we evaluated 18 LLMs, including state-of-the-art models, and observed model performance degradation with increasing input length

[Founders Inc @fdotinc]

We want to invest $100k-$200k in someone ambitious.

If you can take full advantage of this hardware space,

You'll get an invite to a 4 week sprint at our garage:
> on demand access to our team + founders for help.
> stocked lab w/ laser cutters, resin/3D printers, CNC, robot arms, dev kits, soldering stations, + more
> lunch everyday to build.
> work alongside founders just as driven as you.

if you want to change the future, tell us what you're building @ f .inc/apply

[Jeff Weinstein @jeff_weinstein]

AI applications + third party tools is the start of a new form of internet.

[Kent C. Dodds 🏹 @kentcdodds]

MCPs for developers is exciting and cool.

But MCPs for the regular person is where things get much more interesting.

[Ash Martian ✨ @1stMarsColonist]

@karpathy @divya_vikash

[LlamaIndex 🦙 @llama_index]

There are hundreds of official MCP servers out there.

Here’s a nifty tool by @MarcusSchiesser that lets you easily discover, install, configure, and remove new MCP servers from a single CLI interface - supports Claude, @cursor_ai, and @windsurf_ai .

The example shows you how to install LlamaCloud as a server! Ground any popular AI frontend with a knowledge base.

Signup for LlamaCloud: https://t.co/yQGTiRSNvj

[LangChain @LangChain]

🔍 DeepGit: Repo Discovery

DeepGit is an open-source AI system that uncovers valuable GitHub repositories through semantic search. Built with LangGraph, it analyzes code, docs, and community signals to surface hidden gems.

Check it out 🚀
https://t.co/B5etkWXxhi

[Nikita | Scaling Postgres @nikitabase]

There are two flavors of MCP servers: local and remote.

1/ Local MCP servers run alongside Cursor, Windsurf, Claude Desktop, and other MCP clients. You set up the config, give it an API key, and it listens for requests from the client.

Under the hood, a local MCP server typically:
- Runs as a Node.js process (e.g. npx [at]neondatabase/mcp-server-neon)
- Communicates with the AI agent via standard input/output (stdio)
- Uses your locally provided credentials (like a Neon API key) to talk to external APIs

This works well for developers in their local IDE. But ideally, MCP servers should be easy to maintain and accessible to AI agents over the internet, similar to how humans have access to their apps via the App Store.

2/ Remote MCP servers, on the other hand, live in the cloud, connect to the client via Server-Sent Events (SSE) and HTTP, and use OAuth for a secure flow.

What’s great about remote servers:
- No setup required
- Just sign in and go
- Access to the latest tools and updates
- And most importantly, they grant tool access to web-based agents such as Replit, v0, Devin, and others

Remote MCP servers unlock the possibility for AI agents to access a wide variety of tools over the internet. AI agents should "just work," and users shouldn't worry about which MCP server to install.

In the near future, AI agents will be able to infer which MCP server is best suited for the task at hand, connect to it on demand, and take action. However, as agents gain more autonomy, managing access and permissions becomes increasingly important. But that's a topic for another day.

At Neon, we're excited about this trajectory and believe agents will manage infra for you. That's why we released our MCP server back in December 2024, and launched a remote server a few weeks after Anthropic added support for HTTP and SSE.

So if you're a developer, give our MCP a try and let us know how we can improve your experience.

If you're building an agent, DM me to talk about how we can help your business.

[Brendan Irvine-Broque @irvinebroque]

if you've been building MCP servers on Cloudflare (a lot of you have!) — update to get the latest:

npm i --save [email protected]

@JeremyMorrell just saved you money by letting your MCP server hibernate while clients hold open connections to it.

[Sarthak Chauhan @sarthakiitd]

Yes — we’re seeing this too. With large, denormalized schemas, the joins and business rules are rarely explicit. We’ve been exploring ways to encode this latent context — through few-shot examples, query memory, and tool-assisted grounding. Curious how you’re handling schema disambiguation and logic injection — purely via prompt scaffolds, or some structured layer on top? Would love to jam more on this if you’re up for it.

[Sarthak Chauhan @sarthakiitd]

This is fascinating. Been thinking about similar interfaces — is the YC agent framework like Glean or Cursor, with internal tools connected via an MCP-style layer? Or closer to Manus AI / Type & Automate, where natural language drives org logic? Curious how the form factor feels in practice — CLI, chat, something else?

[Sarthak Chauhan @sarthakiitd]

thomasfromcdnjs 5 days ago | parent | prev | next [–]

I've been working on BLAH - Barely Logical Agent Host (https://t.co/wmcNoEkbY8...) for the past few weeks.
It is essentially a standard (has a schema) that has an ecosystem of tools around it.
(completely opensource, no protocol/bridge lockin, no vendor/provider lockin, no ide/client/auton lockin, http/sse/jsonrpc/whatever, local/remote, composable)
So far I'm categorically calling MCP a "bridge", because BLAH supports other bridges such as SLOP (https://t.co/eeQU4kPveQ) and conceptually OpenAPI (or simply HTTP) is a bridge.
An example blah.json looks like this

"tools": [
{
"name": "jsonresume",
"bridge": "mcp",
"command": "npx -y @jsonresume/[email protected]",
},
{
"name": "slop_example",
"description": "Slop example",
"bridge": "slop",
"source": "https://t.co/t831B1y3T9"
},
{
"name": "normal_http_example",
"description": "Normal HTTP example",
"bridge": "openapi",
"source": "https:/example.com/openapi.json"
}
],
}
So blah can orchestra/compose any number of bridges, and it can list any tools it style/format of any bridge. (sorry that sounds so dumb)
For example, you can run `blah mcp start --config blah.json`, and add that to your cursor/claude mcp servers. When you fetch the tools, it loops over all your tools in your blah.json and fetches them whether it is an mcp server, slop server or openapi, it's agnostic and will return a full list of tools pulling from all bridges.
And then you can do `blah slop start`, which will do the same, but the opposite, start a slop server but it will boot up mcp servers and serve them over http too.
So essentially a bridge/protocol agnostic system, you ask for a list of tools, and it traverses everything that can list tools and splats them all together.
That's a little of the main pitch but there are larger plans to try build an ecosystem like npmjs where there is a public registry of any tool (which are just functions at the end of the day). Clouldflare AI team is really getting it right now (https://t.co/db44fUVWcF...), most people would rather just tools being hosted for them, and workers is perfect for that.
It also means you just have to add one MCP server and use one configuration file to manage your tools. (Also has way more control over logging which MCP devs would love)
---
And that all maybe sounds complicated, but it's meant to make things A LOT easier. The https://t.co/X7Tzj4rG7E is horrible, just lots of random ideas and my comment isn't much better, but I'm aiming to have an alpha version in a week or so. Would absolutely love feedback!
4.5 Deep Research Report on BLAH - https://t.co/TrBbi4RWNJ...
reply


haakjacobs 3 days ago | root | parent | next [–]

I agree, people don't want to manually add all MCP servers they want to connect to. I think we're headed in a direction where dynamic tool discovery will be important. E.g. I want to do "x" which requires doing "y" and "z". The client or server should then search for the tools (or agents as tools) that can do the job. So this would basically be a semantic search for which tools are most likely to do the job, then use those tools, and search for more tools if it needs to. RAG for tools, essentially. I wrote an article on how to give ChatGPT unlimited functions a year ago, where I implemented this method.
I basically created a vector database for functions, and created the store based on the tool description, then added the tool definition as metadata. Then for every request I ran a lookup for tools, added them to the tools list. Since tool definitions are added to the context we would have to do something like this at some stage when we want to implement a lot of tools.
https://t.co/wYR32rjBaA...
reply


[Ryan Petersen @typesfast]

If you don't work in global logistics you may have no idea how ground breaking AI-powered natural language report generation is. We will save customers soo much time with the ability to just ask a question to generate a report or graph, then click save to turn it into a dashboard.

[Sarthak Chauhan @sarthakiitd]

https://t.co/HCKdFYNPUt

We introduce SoftTiger, a clinical large language model (CLaM) designed as a foundation model for healthcare workflows. The narrative and unstructured nature of clinical notes is a major obstacle for healthcare intelligentization. We address a critical problem of structuring clinical notes into clinical data, according to international interoperability standards. We collect and annotate data for three subtasks, namely, international patient summary, clinical impression and medical encounter. We then supervised fine-tuned a state-of-the-art LLM using public and credentialed clinical data. The training is orchestrated in a way that the target model can first support basic clinical tasks such as abbreviation expansion and temporal information extraction, and then learn to perform more complex downstream clinical tasks. Moreover, we address several modeling challenges in the healthcare context, e.g., extra long context window. Our blind pairwise evaluation shows that SoftTiger outperforms other popular open-source models and GPT-3.5, comparable to Gemini-pro, with a mild gap from GPT-4. We believe that LLMs may become a step-stone towards healthcare digitalization and democratization. Therefore, we publicly release SoftTiger models at scales of 13 billion and 70 billion parameters, as well as datasets and code for our innovative scalable evaluation, hopefully, making a significant contribution to the healthcare industry.

[Sarthak Chauhan @sarthakiitd]

Check out "Building Ramp's MCP server" by Junho Choi https://t.co/ZeUwuHHszA

[Sarthak Chauhan @sarthakiitd]

https://t.co/qFDRSvZSZO

Future wireless networks aim to deliver high data rates and lower power consumption while ensuring seamless connectivity, necessitating robust optimization. Large language models (LLMs) have been deployed for generalized optimization scenarios. To take advantage of generative AI (GAI) models, we propose retrieval augmented generation (RAG) for multi-sensor wireless environment perception. Utilizing domain-specific prompt engineering, we apply RAG to efficiently harness multimodal data inputs from sensors in a wireless environment. Key pre-processing pipelines including image-to-text conversion, object detection, and distance calculations for multimodal RAG input from multi-sensor data are proposed to obtain a unified vector database crucial for optimizing LLMs in global wireless tasks. Our evaluation, conducted with OpenAI's GPT and Google's Gemini models, demonstrates an 8%, 8%, 10%, 7%, and 12% improvement in relevancy, faithfulness, completeness, similarity, and accuracy, respectively, compared to conventional LLM-based designs. Furthermore, our RAG-based LLM framework with vectorized databases is computationally efficient, providing real-time convergence under latency constraints.

[Sarthak Chauhan @sarthakiitd]

https://t.co/HwGwdAmK83

Vision-Language Models have made significant progress on many perception-focused tasks. However, their progress on reasoning-focused tasks remains limited due to the lack of high-quality and diverse training data. In this work, we aim to address the scarcity of reasoning-focused multimodal datasets. We propose VisualWebInstruct, a novel approach that leverages search engines to create a diverse and high-quality dataset spanning multiple disciplines, including mathematics, physics, finance, and chemistry, etc. Starting with a meticulously selected set of 30,000 seed images, we employ Google Image Search to identify websites containing similar images. We collect and process HTML data from over 700K unique URLs. Through a pipeline of content extraction, filtering, and synthesis, we construct a dataset of approximately 900K question-answer (QA) pairs, with 40% consisting of visual QA pairs and the remaining comprising text-based QA pairs. Models fine-tuned on VisualWebInstruct demonstrate significant performance improvements: (1) fine-tuning on Llava-OV results in 10-20 absolute points improvement across benchmarks, and (2) fine-tuning from MAmmoTH-VL yields a 5 absolute points gain across benchmarks. Our best model, MAmmoTH-VL2, achieves state-of-the-art performance within the 10B parameter class on MMMU-Pro (40.7), MathVerse (42.6), and DynaMath (55.7). These results highlight the effectiveness of our dataset in enhancing the reasoning capabilities of vision-language models for complex multimodal tasks.