It's nearly impossible to prevent LLMs from hallucinating, which creates a significant reliability problem. Enterprise companies think, "Using agents could save me money, but if they do the job wrong, the damage outweighs the benefits." However, there's openness to using agents for non-customer-facing parts and non-critical tasks within the company.
The developers of an e-commerce infrastructure approached us because the format of manufacturer's files doesn't match their e-commerce site's Excel format, and they can't solve it with RPA due to minor differences. They asked if we could perform this data transformation reliably. After two weeks of development, we implemented a reliability layer in our open-source repository. The results were remarkable:
At Upsonic, we use verifier agents and editor agents for this. We didn't expect such high success rates from the agents. I'm surprised by how common these data transformation tasks are. This could be a great vertical agent idea. Btw we use this source (https://arxiv.org/pdf/2501.13946)
I get excited when I see things like this, even if they're simple because I think I can build a business on top of it. However, in complex tasks and real-life cases, it's successful in very few instances. I can't trust its stability. This makes me feel like I've been deceived. I believe agents will be used for tasks that require very little intelligence and constant repetition. It's actually an assistant in situations like this demo. I want to use agents everywhere, but they're not successful in their outputs. GPT-4 is still being used at large scales. I don't know what the situation is with high-level usage of models that do reasoning like o1 through APIs, I haven't tried it. I tried Deepseek, and I encountered stability issues with Deepseek APIs. Besides, R1 doesn't have function calls.
I see no evidence that we're anywhere close to "fire and forget" AI that can be trusted to operate independently. But it feels like every business is centered on not only that being inevitable, but incredibly close if not already here.
Yet AI as a powerful tool utilized by skilled humans is already here, but we can't seem to shake free of this false promise.
I agree. The other day I went to an event. One of the speakers said he had a $2.7 billion exit. Everyone in the room believed it because some brilliant people and high-level authorities in the room believed in him and put him on that stage. Nobody thought otherwise because maybe this could be true. It's not logical to claim otherwise about what someone who might have $2.7 billion says (this really happened). Business, I don't know if there's a theory about this, but they think maybe what people who have received billion-dollar investments say might be true because saying things contrary to what they say doesn't gain anything.
Additionally, I don't want to be misunderstood - Agents have started making significant changes in the workforce right now, and we're even building an agent framework. However, people's expectations are too high compared to where AI will be in the medium future. This turns AI into hype. You will Remember what Sama said about Elon's AGI post.
Hi im Onur, As someone with over five years of experience in Python serialization and the Django framework, I've been heavily involved in the development and maintenance of Tiger, a function store that provides isolated storage, auto-documentation, usage tracking, and version control.
Tiger was born out of the need to solve the difficulties associated with library creation and storage in Python. It serves as a GitHub repository and On-Premises service, storing functions as tools for various large language model (LLM) agent frameworks such as crewAI, LangChain, Autogen, and OpenInterpreter. These AI agents, when integrated with Tiger, become capable of executing codes, searching through stored data, reading Telegram messages, and much more.
Tiger is a two-part system. The first part is a Docker container that provides a dashboard and storage. The second part is a Python client library that I've designed to serialize the functions. This client sends the serialized functions to the Docker container for storage, triggers the auto-documentation process, and keeps usage records.
In the development of Tiger, I've utilized various technologies including Flask, Django, Dill, and Cloudpickle. Notably, I've used the Upsonic Serializer for serializing some functions.
Our goal is to establish a strong community and use case that offers tools for AI agents. This allows all AI agents to easily access modern applications and skills without additional effort. Additionally, companies can easily use Tiger to perform their unique functions. You can provide up-to-date tools for any type of AI agent. Imagine your research agents recommending documents to your team members for reading. This is the high level of usability and standardization we promise.
Tiger can enhance any agent with well-designed tools. Therefore, you can use it to impart search capabilities, the ability to run various programs, and communication skills to all your agents.
As someone with over five years of experience in Python serialization and the Django framework, I've been heavily involved in the development and maintenance of Tiger, a function store that provides isolated storage, auto-documentation, usage tracking, and version control.
Tiger was born out of the need to solve the difficulties associated with library creation and storage in Python. It serves as a GitHub repository and On-Premises service, storing functions as tools for various large language model (LLM) agent frameworks such as crewAI, LangChain, Autogen, and OpenInterpreter. These AI agents, when integrated with Tiger, become capable of executing codes, searching through stored data, reading Telegram messages, and much more.
Tiger is a two-part system. The first part is a Docker container that provides a dashboard and storage. The second part is a Python client library that I've designed to serialize the functions. This client sends the serialized functions to the Docker container for storage, triggers the auto-documentation process, and keeps usage records.
In the development of Tiger, I've utilized various technologies including Flask, Django, Dill, and Cloudpickle. Notably, I've used the Upsonic Serializer for serializing some functions.
Our goal is to establish a strong community and use case that offers tools for AI agents. This allows all AI agents to easily access modern applications and skills without additional effort. Additionally, companies can easily use Tiger to perform their unique functions. You can provide up-to-date tools for any type of AI agent. Imagine your research agents recommending documents to your team members for reading. This is the high level of usability and standardization we promise.
Tiger can enhance any agent with well-designed tools. Therefore, you can use it to impart search capabilities, the ability to run various programs, and communication skills to all your agents.
Hi, i am Onur Ulusoy, a Python enthusiast whose experience spans over 5 years with a special focus on Python Serialization over the past 2 years.
In start we just working on fastest Deployment system in python (<1) but after a pivot in Upsonic is our attempt to the function hub for data teams and machine learning groups by leveraging auto-documentation and auto-dependency. We have designed it to incorporate embedded profiling and offer seamless on-premises installation using Docker.
In the course of its development, we recognized an opportunity to extend the functionalities of our function hub to an LLM agent tool hub. We engineered a custom interface release under the name, "Tiger: Neuralink for your AI agents." This interface, licensed under MIT and community focused.
Its transform the thinking of LLM to real world operations. Just like sending a messsage in WhatsApp or reading emails or writing and running python codes and reading documentation of langchain library. For now we have supports for crewAI, LangChain and AutoGen.
The developers of an e-commerce infrastructure approached us because the format of manufacturer's files doesn't match their e-commerce site's Excel format, and they can't solve it with RPA due to minor differences. They asked if we could perform this data transformation reliably. After two weeks of development, we implemented a reliability layer in our open-source repository. The results were remarkable:
Pre-reliability layer: 28.75% accuracy (23/80 successful transfers)
Post-reliability layer: 98.75% accuracy (79/80 successful transfers)
At Upsonic, we use verifier agents and editor agents for this. We didn't expect such high success rates from the agents. I'm surprised by how common these data transformation tasks are. This could be a great vertical agent idea. Btw we use this source (https://arxiv.org/pdf/2501.13946)