Jason Lee — Associate Delivery Consultant - AI/ML @ AWS

Jason Lee — Associate Delivery Consultant - AI/ML @ AWS · LegalQuants

About

I am a self-taught developer who has spent the last two years working in cloud computing as a Solutions Architect, and now as an AI/ML FDE. I help companies embed AWS technologies to unlock commercial objectives. Prior to this I spent seven years in Compensation/HR across private practice and in-house.

Selected work

Sample Redlining Assistant Using AWSflagship

Published on AWS's AWS Samples GitHub, this repository demonstrates a redlining assistant powered by the Strands Agents SDK and hosted on Amazon Bedrock AgentCore Runtime. It provides an extensible starting point for organisations wishing to build custom redlining assistants, whilst taking advantage of of the scalability, security, monitoring/logging offered by using AWS services. The Strands Agents SDK is a model-agnostic framework created by AWS for building agents. Amazon Bedrock AgentCore Runtime provides the hosting environment for the agent, and can be extended to integrate with other features, such as Amazon Bedrock Knowledge Bases (for RAG). Amazon Bedrock provides the available models. Microsoft JS APIs is used to modify Word documents. Authentication is managed by Amazon Cognito (which allows for integration with SAML compatible identity providers such as Entra ID). Authorisation is managed mostly via AWS IAM, with the exception of inbound invocations to AgentCore Runtime which uses JWTs. This current starting point can be extended by adding features such as memory, MCP targets that integrate with an organisation's AWS environment, added guardrails and fine-grained access control for tools.

Word Add-InOpen Source

Source

Sample Redlining Assistant (Personal Version)

A follow-up side project to my AWS Samples Word Add-In. This assistant lives outside the AWS ecosystem and does not use temporary credentials that AWS IAM roles provide (but rather, long-term API keys that we're all used to purchasing from LLM providers). Here, I experimented building out additional functionality including: - Additional Word modification features - Ability to add MCP servers - Ability to add and switch between LLMs - A UI that looks less AWS-branded - Chat session management (storage, resuming sessions, deleting sessions)

Word Add-InOpen Source

Source

Louis (an AWS-native fork of MikeOSS)

An AWS-native fork of the open-source MikeOSS built to challenge the assumption that law firms must rely on closed, third-party vendors (who themselves are built on the cloud) to achieve scalability and security. Benefits of this architecture include: * Temporary IAM roles vs long-lived API keys * Serverless - no need to manage servers to scale * Observability out of the box - via native monitoring solutions like CloudWatch and CloudTrail to collate application logs and API calls * Extensible - easily switch on additional AWS security (e.g. rate limiting, web application firewall) or other features * Standing on the shoulders of giants - by building directly on the cloud, the app inherits many managed services which are maintained daily by thousands of engineers. * Data perimeter control - data is kept within your own AWS account boundaries vs going to third-party vendors (unless you allow it). For example, invocation requests to Bedrock are not shared with third-party model providers nor routed out of the AWS Region(s) that you choose.

Web AppOpen Source

Source

Clause CoWork

Clause CoWork is an experimental AI-powered legal document analysis tool that leverages a user's existing agent harness (e.g. in the form of an ACP-compatible agent like Claude Code or Kiro) to build a light Karpathy-style wiki for a workspace of documents, and enrich them with metadata for easier navigation and cross-referencing for lawyers. In essence, it is like building a UI wrapper around Claude Code, rather than recreating a harness that is probably more limited in functionality than the ones provided by the LLM providers. The main interface is a web application which operates at a 'workspace' level. There is an accompanying Word Add-In with a more limited set of features. Both read from the same backend SQLite database which populates the two UIs with the clause/document metadata. The agent uses Skills to read and write to this DB. Advantages of using ACP vs creating a harness include: - Using an existing agent harness developed by the LLM provider, and thus being able to access its native toolset (e.g. bash, web search) without you having to reinvent the wheel and implement it yourself - User's get the agent's built-in safety/permissions model (e.g. permission prompts) - Uses the existing agent harness's login mechanism (typically OAuth, which provides temporary credentials rather than long-term API keys) - Can leverage existing agent harness's ability to create subagents, allowing for parallel workstreams to parse larger workspaces - Can leverage one's existing MCP/Skills/AGENTS.md setup Disadvantages: - Less control over the configuration of the harness - e.g. system prompts, tools - ACP is still relatively new - can be challenging to set-up and make 'harness-agnostic'

Web AppOpen Source

Source

1 / 4