Skip to main content

research-lab-canada-az1-ai-entrepreneurs-mike-smith-hal-casteel-2026-01-15-18-58-est-notes-by-gemini

Jan 15, 2026

Canada AZ1 - AI Entrepreneurs Mike Smith/Hal Casteel​

Invited Hal Casteel Mike Smith

Attachments Canada AZ1 - AI Entrepreneurs Mike Smith/Hal Casteel Notes - Canada AZ1 - AI Entrepreneurs Mike Smith/Hal Casteel

Meeting records Transcript Recording

Summary​

Hal Casteel opened the meeting, introducing the platform CODATECT, which is designed with agentic workflows and a unique memory management solution to minimize hallucination and address regulation and compliance, especially in regulated industries like insurance, as noted by Jhared Smith. The participants, including Mike Smith, Jhared Smith, David Q Chen, Mike Wicks (Ross), Abhi Pushparaj, and Attila, discussed their backgrounds, individual AI tool usage, and the challenges of AI governance and memory in large language models. The conversation centered on the importance of a risk management framework, detailed specification in AI development, and utilizing CODATECT's capabilities for developing repeatable, auditable, and high-value industry solutions in planning, scheduling, and supply chain management.

Details​

  • Meeting Opening and Introductions Hal Casteel opened the meeting by commenting on the snowy weather reported by Mike Smith in Canada, contrasting it with Hal Casteel's 25°C weather in Brazil (00:00:00). Mike Smith introduced his son, Jhared Smith, who is located in southwestern Ontario, near Lake Huron (00:03:04). Jhared Smith, an insurance broker by profession, shared that he had previously worked as an accountant for five years before switching to insurance. Hal Casteel also noted that Jhared Smith is an IT guy "through osmosis" of his father (00:04:51).

  • AI Regulation in the Insurance Industry Jhared Smith mentioned that there is a lot of regulation concerning AI in the insurance industry (00:04:51). He also shared that they are scheduled for a regulatory "campfire chat" in February regarding how AI can or cannot be used, as they are currently instructed to avoid using it for the time being (00:05:41) (00:43:28). Hal Casteel, who has a four-decade background in healthcare and worked at Grail, noted the regulation in the insurance business and his own experience in regulated industries (00:05:41) (00:10:54).

  • Introductions of Additional Participants David Q Chen joined the meeting; Hal Casteel and Mike Smith helped troubleshoot an audio issue with them (00:06:38). Hal Casteel noted that they previously met David Q Chen when they were the CEO of Benchai, a startup co-founded with Abhi Pushparaj (00:07:52). David Q Chen is now semi-retired and taking a break (00:08:53) (00:14:30). Mike Wicks, who joined later, was using a company email, and shared that their background is mainly in the sales side of ERP solutions (00:08:53) (00:15:34). Abhi Pushparaj joined and apologized for the delay due to dealing with their baby, and Hal Casteel proceeded with the round-robin introduction (00:14:30).

  • Individual Backgrounds and AI Experience Mike Smith, a Solution Consultant at Oracle Netsuite, shared that they work with Hal Casteel and are interested in learning more about AI development in the ERP space (00:10:00). Jhared Smith's background is in finance and insurance, and they became interested in AI after seeing it shrink the accounting field (00:13:38). Abhi Pushparaj previously co-founded an AI search engine company on antibodies with David Q Chen (00:14:30). David Q Chen has a background in neuroscience and is currently looking for the next venture, including exploring neuroscience research related to a compound called "ebook game" with Abhi Pushparaj. Mike Wicks, who prefers to be called Ross, confirmed that they had been on an initial call with Hal Casteel and Mike Smith, and they are currently working for a smaller partner focusing on data and business intelligence tools (00:15:34).

  • Personal Use of AI Tools When discussing personal AI use, Jhared Smith noted they predominantly use ChatGPT for quick research and are dabbling in other tools recommended by their father (00:17:33). David Q Chen uses cloud code with Cursor and also Grok for research, noting they sometimes use Cursor and Claude. Abhi Pushparaj primarily uses Grok and Claude. Hal Casteel asked about Perplexity for research, to which Mike Wicks responded that they use it for tweaking messages, finding it better than ChatGPT (00:18:29). David Q Chen, who subscribes to Grok, relies mostly on that tool for research (00:19:36).

  • Agentic Workflows and CODATECT Platform Hal Casteel inquired if David Q Chen or Abhi Pushparaj were using agentic workflows; David Q Chen had experimented with them but had not gotten deep into it, and Abhi Pushparaj had not yet started (00:19:36). Hal Casteel introduced their company, awan.ai, and their platform, CODATECT, which they intend to demo and provide access to selected individuals for pilot feedback (00:12:20) (00:20:47). Hal Casteel noted that CODATECT, built on Claude code, uses a combination of agents, skills (similar to prompts), and scripts (Python, JSON, shell scripts) to direct the LLM and minimize hallucination (00:23:02).

  • CODATECT Development and Features David Q Chen asked whether Hal Casteel or the Claude-foster agent wrote the code for CODATECT. Hal Casteel explained that development started in March of the previous year, and CODATECT, now in its seventh version, has largely built itself on top of Claude code since August 27th (00:28:00). A unique feature of CODATECT is its ability to export and manage memory, allowing for running multiple sessions in parallel (00:29:19).

  • Introduction of Attila Attila joined the meeting and was introduced by Mike Smith as a riding partner. Attila shared their background as an avid cyclist running a cycling tour business and being deeply involved with ERP implementations for logistics and warehouse operations (00:30:42). Attila noted their strong interest in AI and how they have fully switched their marketing and clothing line design to using AI tools (00:32:17).

  • Regulatory Focus of CODATECT Hal Casteel explained that CODATECT was initially built to address regulation and compliance concerns in heavily regulated industries, such as food distribution, medical devices, and insurance (00:32:17). The platform incorporates tools like a risk management framework, multi-user, and multi-tenant capabilities, allowing multiple businesses and teams to manage projects in parallel (00:33:27). Hal Casteel highlighted that the system allows for tracking changes and procurement, which is relevant for FDA regulations for medical devices (00:34:50).

  • FDA Regulation of AI in Medical Devices Abhi Pushparaj confirmed that FDA regulation on incorporating machine learning or AI into devices is still evolving and currently requires freezing developments before commercial release (00:34:50). Hal Casteel's framework is built on NIST and GDPR requirements and is designed to evolve as regulations firm up, with a mandate for more firm guidance from the FDA this year (00:36:06).

  • Risk Management Framework for AI Adoption Hal Casteel detailed the importance of a risk management framework, particularly for governance, to manage AI usage across an organization. They suggested that both new and existing businesses should implement such a framework early on, which includes defining a charter for safe and responsible AI adoption, assessing risk, and documenting policies and processes (00:39:38). Attila found the framework intriguing because they are facing issues at their company with employees misusing AI and providing incorrect regulatory information to bodies like the CFIA (Canada Food Inspection Agency) (00:51:51).

  • Communicating AI Risk to Leadership Hal Casteel emphasized the need to communicate AI risks effectively to management, including executive summaries and board presentations to ensure leadership understanding and alignment (00:42:05). Hal Casteel encouraged adopting a proactive approach to AI adoption to avoid competitors gaining an advantage and stressed the importance of not using unapproved personal accounts or browser extensions (00:43:28). Mike Wicks supported this, noting that their attempts to target law firms with AI projects are often halted by the IT department due to the lack of documented guardrails and risk information (00:46:28).

  • Challenges and Risks in AI Hal Casteel shared that the development of CODATECT was motivated by the lack of an existing risk management framework at Grail (00:46:28). David Q Chen highlighted their concerns about long-term memory problems in AI, where models often lose context as projects grow, leading to hallucinations and output unreliability (00:50:39). Hal Casteel addressed the concept of an "AI back," noting the hype cycle has peaked, with 95% of internally developed AI projects failing, leading to a "trough of disappointment" where smart investments based on lessons learned should occur (00:54:43). Hal Casteel added that projects managed by third parties are yielding higher returns (00:55:52).

  • Analyzing Work and Project Management Hal Casteel discussed the importance of analyzing work and workflows to determine where AI should be appropriately applied, noting that automation is not inherently dangerous. They introduced the concept of work helix, which focuses on breaking down work activities to find opportunities for automation (00:57:16). Hal Casteel suggested that changes in an organization, like software projects, should be managed with project plans, breaking down efforts into smaller components (00:58:37).

  • Importance of Intention and Specification in AI Development Hal Casteel emphasized that in working with AI systems, especially large language models (LLMs), intention is crucial and requires detailed, unambiguous documentation because ambiguity is "like poison" to a generative LLM (00:59:51). Hal Casteel suggested spending about 50% of the development time on specification to ensure the intention is thoroughly documented, which leads to higher quality outcomes. David Q Chen agreed, noting that most of the work with AI agents is in defining the task and organizing the definition and evolving context (01:01:29).

  • Ambiguity in Human Language and LLM Hallucination Hal Casteel noted that natural human languages are full of ambiguity, which historically reduces conflict among people, but ambiguity is dangerous when interacting with LLMs. They explained that hallucination is a direct consequence of ambiguity, detailing that natural language is translated into mathematical tokens for the LLM, and if those tokens are not explicit to map the intention clearly, the LLM relies on statistical evaluation, potentially leading the response off track (01:02:34).

  • The Concept of Context and Memory in LLMs Hal Casteel equated context with memory, illustrating that individual experiences and points of view lead to different memories of the same event, which affects the output of LLMs based on individual input (01:04:00). David Q Chen raised a concern about trusting the LLM to perform as intended and needing to constantly read the code to prevent problems downstream, noting that they read a lot of code but do not write as much (01:06:23).

  • Agentic Systems and the Quality-Quantity Challenge Hal Casteel discussed the challenge of managing the massive output generated by multi-agentic systems, noting that output from many agents across five sessions is beyond human ability to evaluate, and quantity does not equal quality. Hal Casteel introduced the concept of a "mixture of experts" (Mixture of Experts) in agentic systems, where various roles like system architect, product manager, programmer, and tester are represented by agents, and a separate Mixtue of Experts of "judges" conducts qualitative assessments to ensure work meets defined standards (01:08:35).

  • Memory Management and Context Compaction Hal Casteel identified memory and context management as the most critical problem to solve in Gen AI systems, noting that LLMs have a limited context window (a "bucket of tokens") (01:11:20). Hal Casteel explained that Anthropic's system, for example, will auto-compact its context arbitrarily when it exceeds about 76% of its capacity, collapsing the information it decides to remember for the next session (01:12:44).

  • The Code of Tech Memory Solution Hal Casteel described a solution developed with CODATECT to manage memory by defining small work units (tasks) (01:14:12) and exporting the entire context at 75% capacity into a JSON NL file. Hal Casteel created a script to parse these files into individual messages and load them into a growing SQLite database (01:15:29). This approach allows agents to query the database and selectively retrieve only the "relevant context" pertinent to their current work unit, which Hal Casteel sees as a significant evolution (01:16:47).

  • Long-Term Memory and Future Plans for CODATECT Hal Casteel noted that their system provides "real long-term memory" by capturing context into the smallest possible unit associated with any particular activity (01:18:52). Hal Casteel aims to make CODATECT a multi-tenant platform for enterprises dealing with massive and complex data sets, particularly in healthcare and finance (01:19:56). They also mentioned being interested in the concept of a "recursive language model" and applying scientific research into CODATECT, which already performs scientific research well (01:21:01).

  • Building a Collaborative Platform and Ecosystem Hal Casteel expressed a desire to work with collaborators who will use CODATECT as a platform to build new solutions and potentially businesses within the framework, emphasizing that CODATECT will not own what others create (01:22:10). Hal Casteel acknowledged that businesses in various sectors, including insurance, are facing similar problems that CODATECT can help address by defining overarching rules like legal frameworks, enterprise policies, and standard operating procedures (01:23:14).

  • Human Value and Future Workflows with Agents Hal Casteel suggested that AI is beginning to dissolve teams because individuals can accomplish much more on their own, but they seek to keep humans in the loop while leveraging the power of a team's diverse points of view to solve problems (01:24:31). Hal Casteel noted that the system allows users to automate the export of source files, which includes documents with a standardized structure for database indexing (01:27:14).

  • The Context Extraction Process and Local System Functionality Hal Casteel demonstrated the context extraction process, showing how the CX command runs a Python script to parse session files into databasable units and load them into the database, generating hundreds of thousands of messages since early December (01:29:54). Hal Casteel explained that a local Rust binary watches the token count and automatically exports and indexes the context at 75%, allowing agents to directly query the database for relevant context independently of the user (01:31:20).

  • Inputting Data and Analyzing Artifacts Hal Casteel demonstrated how input can be given to CODATECT through various artifacts, such as dropping files like ideas and images into a dedicated folder, or via direct prompting using a CR command (01:35:34). Hal Casteel utilizes the platform to automate the export and analysis of YouTube transcripts, converting them into text files that are processed by a set of agents and skills (01:32:48). Hal Casteel also mentioned building a self-developed, cheaper containerization platform similar to Modal, deployed on Google cloud workstations for multi-tenant and multi-user environments (01:34:09).

  • Enterprise Integration and Extensible Architecture Hal Casteel clarified that CODATECT's local systems are pushed to a PostgreSQL database in the cloud, with user management and access control systems built on top (01:36:48). Hal Casteel confirmed that CODATECT could write solutions that interface with systems like Microsoft Fabric through APIs, allowing users to run CODATECT on their own and extend the platform (01:38:04). Hal Casteel described the platform as an extensible "kit of parts" where users can build their own patterns and components, highlighting that they do not want to use platforms that dictate how they must work (01:39:03).

  • Open Source Decision and the "Witch" Command Hal Casteel decided to keep CODATECT closed source initially, believing that releasing it now could "fuel massive unemployment" and that society is not yet ready for that level of disruption (01:40:51). Hal Casteel demonstrated the "witch" command, which searches the database of agents, skills, and components to recommend the most appropriate mixture of expert agents for a given task, such as assessing the quality of a risk management framework (01:42:02). Hal Casteel confirmed that the system is capable of completely automating task planning, sequencing agents, and execution (01:43:23).

  • Agent Orchestration and Future Business Workflows Hal Casteel demonstrated how the system proposes multiple options for task execution, such as using a Judge Panel for compliance-related work, and that users can string these agents and workflows together like programs (01:44:29). Hal Casteel emphasized the need to think about what business workflows will look like in the future with agentic systems, noting they may not need to fully follow human workflows (01:45:49). Hal Casteel estimated the value of one developer running on CODATECT to be around a $6 million annual return in terms of replacement value (01:47:01).

  • Horizontal Solutions and Service Business Opportunities Hal Casteel described CODATECT core as the brain of operations, supported by 84 other components, including a workflow analyzer for enterprises designed to data mine enterprise workflows and build "horizontal solutions that can work across multiple verticals" (01:48:17). Mike Smith suggested using the platform to create a service business for small to midsize businesses (SMBs) that lack the resources of large consulting firms (01:49:24) (01:52:03). Hal Casteel sees CODATECT as a potential Work Helix replacement for the SMB market (01:50:51).

  • Business Development and Service Partnerships Hal Casteel shared that during their business accelerator experience, 15 businesses offered them equity and checks to build their products, which led Hal Casteel to partner with seven development shops to bring them up to speed on CODATECT to manage the demand and build parallel service arms (01:52:03). Hal Casteel is also building a CRM into the platform for lead generation and prospect management (01:54:55). The group agreed that service and consulting based on CODATECT's capabilities, applied to areas like insurance workflows, presents a huge opportunity (01:55:48).

  • Developing High-Value Industry Solutions Hal Casteel proposed creating a requirement and working together to fully shape out a high-value industry solution that CODATECT can build quickly. Mike Wicks noted a need for AI tools in planning and scheduling, especially in discrete, batch, and process manufacturing, to predict needs with constantly changing market ingredients (01:57:48). Mike Smith and Hal Casteel agreed that there are opportunities in planning, budgeting, forecasting, and supply chain management to address existing gaps (01:59:13).

  • CODATECT Platform Architecture and AI Strategy Hal Casteel differentiated CODATECT as both a platform for creation, modeling, and measurement, and a tool for design and architecture. Hal Casteel advocated for using LLMs only where appropriate, suggesting that people may be overusing them and instead recommending a neuro-symbolic programmatic approach to break down problems into cheaper, manageable program units that save tokens and energy (02:00:26). Hal Casteel stressed that for business problems, systems need to be built to be repeatable, reproducible, explainable, and auditable, which LLMs alone may not guarantee (02:02:29).

  • AI Application in Business Functions Mike Wicks observed that while AI projects often start with basic HR co-pilot solutions, they have not widely seen AI used in planning and scheduling, unlike reporting and forecasting, which are big priorities for most companies (02:02:29). Mike Smith described using AI for project analysis to predict critical mass and resource needs, which is particularly valuable for project-based companies or service businesses (02:03:41). Mike Wicks shared a case study of a friend's business that builds transformers for Med Device Aerospace, which could use historical ERP data and previous quotes to quickly generate estimates and maintain competitiveness, potentially utilizing a Configure Price Quote (CPQ) functionality (02:04:37) (02:06:49).

  • Dynamic User Interfaces and Data Mining Hal Casteel discussed integrating CODATECT with open-source 3D modeling tools to rapidly generate shapes on demand and enabling rapid iteration with AI (02:05:41). Hal Casteel also introduced the idea of dynamic user interfaces, where UIs change based on data interaction, suggesting that they don't need to be hardcoded and can be built using preconfigured React objects that self-assemble based on the use case. Hal Casteel mentioned the platform's ability to mine data from spreadsheets and properly database it, which is useful for building UIs on top of existing data (02:11:38) (02:13:18).

  • Proposal for Increased Meeting Cadence and Strategic Focus Mike Smith proposed increasing the meeting cadence to a bi-weekly schedule, which Hal Casteel and Dunnville Grand Tour agreed to (02:13:18). Hal Casteel emphasized the goal of cherry-picking high-value ideas that are reproducible and scalable products, aiming to build a product suite rather than focusing on single one-off customers, and expressed openness to white-labeling the built solutions (02:14:06). Hal Casteel confirmed interest in a pilot run with the owner of Elletron as a case study for Med Device Aerospace, suggesting that regulatory projects are a long-term vision, but the current focus should be on faster execution (02:08:42) (02:15:32).

  • Future Collaboration and Development Support Hal Casteel offered to start working immediately with Mike Smith to flesh out requirements and also invited anyone with B2B or even consumer ideas to reach out for assistance in development, citing an example of successfully building a consumer app called "Hookup" (02:14:06) (02:17:22). Mike Wicks requested assistance on connecting CODATECT to Microsoft solutions, as their network is heavily invested in that platform. Mike Smith suggested that the next session could focus on the setup and getting CODATECT up and running, including discussing the availability of install scripts, Docker, and Google cloud workstations for different operating systems (02:24:28).

Suggested next steps​

  • Hal Casteel will share the CODATECT risk management framework freely with the participants of the meeting so they can be more informed and use it as a starting point.
  • Hal Casteel will do a transcript of the session.
  • Hal Casteel and David Q Chen will talk more on a technical level to look at what Hal Casteel's parser is doing to see if they can refine the parser to capture the context into the smallest possible unit.
  • Mike Smith and Hal Casteel will talk about how they should be thinking about the different ways of modifying the business consulting opportunity.
  • Mike Smith will work with Hal Casteel over the next two weeks to prepare for the next session, which will cover the setup and getting Codec up and running.
  • Mike Wicks will try to invite Pu to the next call to discuss a basic solution for his needs and potentially pitch solutions to his competitors.
  • Mike Smith will send out the bi-weekly meeting invitations through his own account to ensure the collective is kept together.
  • The group will increase the meeting cadence to a bi-weekly meeting.

You should review Gemini's notes to make sure they're accurate. Get tips and learn how Gemini takes notes

Please provide feedback about using Gemini to take notes in a short survey.