top of page
Engineer

Add Intelligence with
AI Agents

We orchestrate language models, such as OpenAI's GPT4, into agents who complete tasks.

These agents can be arranged into simple automations, complex workflows or even creative conversations. Agentic AI to automate tasks is our specialisation.

"The set of tasks that AI can do will expand dramatically because of agentic workflows"

Andrew Ng, founder of Google Brain, Baidu Chief Scientist, Director of Stanford's AI Lab

March 2024

Market Intelligence

  • Market & Competitor analysis

  • Lead generation and qualification

Knowledge Management

  • AI-powered search & summarisation

  • Knowledge base creation

Customer Experience

  • Sentiment analysis

  • Chatbots for sales + service

Process & Data

  • Automated code writing + testing

  • Automated data labelling

The Process

Slide7_edited.jpg

01
Identify Automation Opportunities

  • What are the goals, business model changes, staff expectations, technical goals?

  • Our Approach: Discover business context from management. We interview staff (demand led) and make suggestions (supply led) for automation opportunities in the business

Studying in a library

Case Study
Market Research

Problem

The client wanted to automate a time consuming web research task. They would monthly search for competitor products, discovering the latest features and news, then boil that down into a spider graph comparing each product on common features. The benefit of AI agents is that this research be automated and updated on a monthly basis.

Solution

We arranged large language models (LLM), like ChatGPT, into AI 'agents' to allow them to collaborate in a team. Each agent was be given a specific task within a workflow, just as a person would
 

The AI agent's workflow is similar to a person's for this task:

  • Online web search for product information

  • If relevant information is found then the agent extracts it, else it continues the search until some limit is reached

  • Collate and summarise product information

  • Critique the summary versus sources for accuracy

  • Score the product against given criteria, or propose criteria

  • Present all products on one table or chart

Recipe

The recipe for an agent is = LLM + Data + Tools + Environment

Recipe Continued...

- LLM

We used Anthropic's Claude 3 Opus for reasoning, such as comparing products, and Anthropic's Sonnet for summarising long texts. Other models were made available, OpenAI GPT4 for reasoning and GPT3.5 for summarising.

- Data

The data comes from web searches via a service called Tavily, which allows AI agents to search the web in general. We then custom built a bot to explore deeper within a discovered website for the product information.

- Tools

- In addition to web searches, the AI team will have the ability to save their results in a tabular format, we chose excel and markdown.

- The team use a 'machine learning' tool called text embeddings to plot the products on a spider graph, comparing them on common features.

- Environment

We grant the agent team access to the internet and a safe file location for saving their results.

Otherwise, all the team simply run on the client's local server , making calls to the LLM (Calude3) as required, as this is publicly available information.

Digital Programmer

Case Study
Data Science

Problem

To fully automate the work of a data scientist with AI agents, expanding productivity of human data scientists by exploring options they had no time to consider. AI agents had to be adaptive to any situation, no step by step process could be prescribed. 

 

Solution

This was a research project by Agentico for Microsoft Research and was so successful that is was built into their AutoGen project. Agentico built AI agents as data scientists, exploring data and algorithms given any objective. Not only are they highly adaptable, but human data scientists can easily inspect, repeat and adapt the agent's work in the same environment as the team were working. 

Recipe

The recipe for an agent is = LLM + Data + Tools + Environment

- LLM

The AI is asked to take on three roles:

  1. A data scientist / software coder

  2. A critic to review the coder's work

  3. A conversation manager, intelligently choosing which of the four participants should speak or act next

Recipe Continued...

 

- LLM

A fourth role is played by straightforward code, not an LLM :

4: The coding environment, safely executing code and returning results or errors for the team to review or correct.

These roles were all played by OpenAI's GPT4, which remains the best at writing software. Role 3 is a foundational feature of Microsoft's Autogen, it is core to achieving solutions which have weakly described processes, or none at all.

- Data

With a flexible objective, no single data source can be provided. Instead, the team are enabled to open a folder and find files of data via code.

- Tools

- The team can execute code, this allows any tools which can be accessed via code, the internet or an API. 

- Environment

The bulk of the work was in safely providing the team of agents with access to an environment, in this case Jupyter Notebooks. The notebook also records the team's conversation (i.e. reasoning), code and code results, a useful audit trail of the team's activity.

Watch the team in action...

bottom of page