Ajay Krishnan
@Aj_Krish23
Followers
92
Following
1K
Media
3
Statuses
52
LLMs, Engineering and a ton of music Curr: DevRel @UnstructuredIO Prev: SWE-Intern @intel, MLE @MadStreetDen, Cool Research @SpaceMLorg
New York
Joined May 2020
So, a brain?.
Imagine running a local llm inside of your neuralink implant. This is where we are heading.
0
0
2
Reimplementing backpropagation from scratch has been humbling. Crazy how much we take autograd for granted.
0
0
1
LangGraph has been one of the easiest/most customizable tools I've used recently to build Agentic Workflows. Every workflow is a graph, each node is a function call. Make a couple nodes LLMs and you have an Agentic workflow all set up.
🧠Agentic RAG with LangGraph and @togethercompute🔍. Explore how to build an Agentic RAG system for understanding startup equity that can reframe questions and dynamically retrieve better information using LangGraph and models from TogetherAI. Unstructured API provides the.
0
1
3
RT @UnstructuredIO: 🛠️ Want to build a complete ETL pipeline for your unstructured data?. Join us next Tuesday 4/29 for a hands-on walkthro….
0
3
0
Built this with MCP, Unstructured and @databricks to start off DocProc workflows with plaintext!.Go through the setup instructions, provide Claude what it needs and you get workflows up and running in minutes🙂.
Imagine building a full ETL pipeline with a single sentence—no code, just conversation. With Claude Desktop, Unstructured, and Databricks all wired together via Model Context Protocol (MCP), you can move data from S3 to Delta Tables using only natural language. We walk through.
0
2
5
RT @METR_Evals: We observed o3 in particular has a propensity to try to “hack” our tasks to get a higher score. Importantly, we saw this ar….
0
20
0
Damn.
o3 overall feels like a system -- i loved how it analysed the images - note that it is literally thinking in images sort of because all this is still happening in the CoT stage
0
0
1
Yeah I'm reinventing the wheel w o1 in the picture, but I can atleast track it's thoughts and tune them around how I want them to go.
0
0
0
My favorite recent GPT-4o hack for non coding work lately:.Forcing it to loop its own thoughts. " start a <think> section and reflect on how to improve it. You can keep typing in <think>. You can run multiple think loops". GPT becomes its own critic. Output now feels better?
1
0
4
25yo me w new LLMs and their increasing performance.
0
0
1
More otw!!!.
🚀New Notebook! . Learn to create a complete data processing workflow with Unstructured API and its Python SDK to transform all of your unstructured data (PDFs, emails, powerpoints and more) into structured data in a Delta Table in @databricks. Why build entire workflows with.
0
0
2
RT @techczech: @Ahmad_Al_Dahle You really need a better cannonical chatbot than - it's just not good enough to show….
0
2
0
RT @UnstructuredIO: 💻 We've got some CODE EXAMPLES to help you get started with the #UnstructuredMCP Hackathon!. Check out this compact MCP….
0
2
0
Lets gooo first of many from last week! Lmk how you end up playing w this 🔥.
🚀 @firecrawl_dev to @DataStax AstraDB with Unstructured MCP! . Using Unstructured MCP Server's integration with Firecrawl, crawl websites, process the data, and load it into AstraDB for RAG applications—all through natural language commands. Learn how:.
0
0
1
@UnstructuredIO Back! @UnstructuredIO x @firecrawl_dev to crawl and crunch documentations and for quick answers and work. Will sharpen this better tomorrow but really excited about how this is shaping up! 🚀
0
4
6
@UnstructuredIO As someone that's also built Docparsers by myself previously, it feels great to be able to blindly trust that the extraction is clean and works out how I want it to and go about my downstream work. But anyways, brb might think of some improvments to what I've built so far
1
0
1
The LLM took my local files, threw them into S3, spawned an @UnstructuredIO Platform Workflow to chunk and ingest text, tables and images and set me up to start playing w RAG all under a minute!.
1
0
2
Spent some hours coding it up .Here's what I got out of it :) .
Go team! Reporting live from @Aj_Krish23's workstation:. MCP, Unstructured Platform API, and Astra DB—got my docs transformed, chunked, ingested, and ready for some RAG all over a quick chat!.Just ask, and boom—data where you need it. Extending to more vector DBs over the week 🚀
1
0
0
A level deeper:. Servers: Expose tools & capabilities through MCP. Hosts: Manage connections to MCP servers. Clients: Connect and execute. In the near future - I hope there's no more duct-taping integrations? And ideally a more accessible way to get work done?.
1
0
1
Highlevel:."What if you didn’t have to deal with integration headaches? We handle the integrations, you just tell your LLMs which tools to call and when. That’s MCP.".
1
0
1