Keywords: LangChain, SpaCy, HuggingFace, NLTK, Llama-Index

• Building effective question answering systems based on company history and knowledge.
• Building a question answering system based on in-house technical reports and customer interaction logs
• Implementing memory management and safeguards for AI assistants
• Building AI assistants with persistent memory
• Developing advanced interfaces in ChatGPT style
• Developing features such as prompt saving and sharing

Keywords: Vision Language Model, Autonomous Agents

• Developing of new autonomous agents
• Automate market research, literature reviews, etc. by customizing AutoGPT, babyAGI, etc.
• Implementing multimodal processing with Multimodal Embedding.
• Implementing a chatbot that allows to input images using embedded vectors.
• Adding chat interfaces to existing applications
• Adding natural language chat interfaces to existing applications

Keywords: Local LLM

• Implementation and tuning of local LLMs running in on-premises environments
• Developing chatbot systems that can be used without exposing confidential data
• Development of a real-time minute-taking system running in on-premise environments
• Generating minutes from meeting audio using local speech recognition models
• Anonymization of unstructured data based on named entity recognition
• Anonymization of names of people, places and addresses in unstructured data to make AI development outsourcing more efficient

Keywords: Ray, Kubernetes, TensorFlow, JAX, PyTorch

• Acceleration of Transformer mechanisms
• Development of large-scale ETL processes
• Distributed training of LLMs
• Visualization of attention between tokens within the language model
• Fine-tuning large language models for custom domains

Question Answering from Large-Scale Unstructured Data using LLM

We used a large-scale language model to build a system of question-answering on large unstructured data without fine-tuning and evaluated its accuracy.

A bottleneck in the application of large-scale language models is the limited number of input tokens. By storing documents in a vector database, which can then be explored and added to the context as required, this problem can be avoided.

Development Environment & Technical Fields: GPT-4, LangChain, Llama-Index, in-context learning, embedding, Semantic Search

LLM recursive automated planning systems

We have built an automated planning system for software development by recursively refining the output of a large language model.

Given a software requirement specification as input, this system interactively checks the details of the requirement on the GUI and decomposes it into the tasks required to achieve the goal. The system is actually used in our company.

Development Environment & Technical Fields: Large Language Models, Autonomous Agents, Token Compression, Prompt Engineering