Accelerating Startups to Put Enterprise GenAI Multi-Agent Systems in Production

EVENT LOGISTICS:
The event is hybrid. The audience can attend in Mountain View, CA or join on Zoom or YouTube

Valley Research Park (VRP)
319 North Bernardo Ave
Mountain View, CA 94043
From the locked front door, go to the right side of the building. The 2nd door will be open.

If you want to join the discussion remotely, you can submit questions via the Zoom “discussions”. The zoom link:
https://acm-org.zoom.us/j/95883310048?pwd=AuzloSgX3UwaDlhhNw2iHg8NPjgzRZ.1
Join via YouTube:
(to be updated closer to the event date)

Agenda (one hour earlier this time, for audience in other time zones)
5:30 people start to arrive, enjoy pizza and networking. Test your connection and chat.
6:00 SFBayACM introduces upcoming events and the speaker
6:00 speaker presentation starts
7:15 - 7:30 finish, depending on Q&A

TALK DESCRIPTION:
– PROBLEM –
Key Challenges We Will Address for Different Audience Segments
1. Startups with Funding, or Looking for Execution Partners:
You have a strong vertical application and early investment but now you need to grow quickly while focusing on your core value proposition. You may already have a prototype built from initial experimentation but face the challenge of evolving it into a secure, multi-tenant, enterprise grade GenAI platform that is compliant, transparent, explainable, and aligned with human values.

2. Mid Cap and Large Enterprises:
Your teams have built GenAI prototypes yet struggle to solve real problems or get applications into production consistently. You want to identify and overcome the technical, operational, and organizational barriers that keep innovation from scaling.

3. Investors, VC’s and Venture Partners:
You are assessing how to reduce costs and risks in your portfolio companies’ GenAI strategies. Understanding why 95 percent of projects fail, as reported in MIT’s State of AI in Business 2025, can help you guide founders toward sustainable and successful models.

4. Product and Engineering Leaders:
You have ideas, and aim to design applications that surprise and delight customers by building intelligence that anticipates user intent, with a Do What I Mean (DWIM) design principle that separates great products from the rest.

5. Regulated Industry Executives:
In sectors such as financial services, healthcare, or government, you must ensure AI systems are reliable, explainable, auditable, and fully compliant with evolving laws and governance standards while maintaining speed and innovation. You may also want to monitor your systems, to comply with various company or security policies.

Who Should Attend:
Startup founders, investors, product leaders, and enterprise innovators who want to understand what it really takes to design, build and scale GenAI systems that work in the real world.

– SOLUTION –-
This talk will answer problems listed above and more. The speaker will first dive into the MIT report, to better interpret the context of the survey. Lessons learned on how to help GenAI deployments succeed will be discussed.

Studies that review knowledge worker productivity gains will be reviewed.

The speaker will discuss an architecture for a flexible software accelerator framework that can meet the needs for successful GenAI deployments. The framework can be configured to support different applications to meet specific requirements, combining best-practice software components and design principles.

Design Principles

• Designed for augmented human intelligence, multi-agents, AI transparency, explainability, alignment with human values, and security.

Differentiators and Innovation

1. In the context of an application, communicate between multiple people and agents over time, building up a medium-term memory, which is at the level of an employee work-task, like a shared canvas with collaboration.
2. The Design of Experiments (DoE) module helps to optimize our framework for your problem, data, requirements and KPI's. You can optimize for accuracy, response time, cost or other metrics over many architectural design decisions.
3. Auto tagging with meaning-markers, to support 1,000's or millions of problem specific products, places, people or things. This enables creating structured queries, graph RAG and better application of guardrail polices with Named Entity Recognition (NER). This is also a way of reliably scaling up n-shot learning, from 5-10 examples to 1,000's, to understand all 6,500 public stock company names and tickers (vs. all bonds or mutual funds), all medical symptoms, diagnosis and treatments. A given tag can be in one or more hierarchies.
4. Built in defense against LLM Scheming. Scheming is when a GenAI system may be exposed to information that may lead it to change its goals, decide to lie to end users, sandbag or act insecure. See past slides or a video on defense mechanisms, which can use Guardiance to monitor agent thought traces, as one defense. Other defenses are architectural, or detect behavior drift over time from the original assigned goals.

Software Architecture
Given an accelerator software architecture with the layers listed in the following sections (accelerator, tools, GenAI tools).
(for a better graphic, see: https://www.ccube.com/luminai, "GenAI Architecture Layers")

Ccube’s Lumin Lab Accelerator software

• IntelliConnect (to connect you to the intelligence of the text and data in your company) This lets users ask questions and reason about of your corporate knowledge. This is a RAG system that can support Vector Databases.
• Guardiance (Guardrails to support compliance and AI alignment.) To simplify, general text classification with many possible types of alert management or different actions.

Tools and software:

• Microservices architecture
• MCP integrations
• Many file types and data sources (i.e. local device drive, Google drive, Gmail, MongoDB, ...)
• Enterprise applications (i.e. SalesForce, Service Now, ...)
• application personalization, knowing the user over time
• workflow support (n8n, UiPath)

GenAI Tools:

• Retrieval Augmented Generation (RAG) with Vector Databases, such as Weaviate or Pinecone
• Use [Guardrails.ai](http://guardrails.ai) as a foundation for text classification. We make use of GenAI or Python to match conditions, +/- examples per condition, and allow a general Boolean combination of conditions into rules. Support many rule actions and rule testing. Support observability of GenAI behavior, to be in alignment with your company values and brand. Applying rules is multi-threaded.
• Multi-Agent with: LangGraph, LangSmith
• Agent logging channels: a) system level (tokens, LLM calls, costs, latency), b) user-agent level chat, c) agent & LLM thought traces. Guardiance can apply rules to text sources, including: email, docs, folders, chats or agent thought traces for scheming detection.

– RESULT –
The speaker will show a series of demos relating to financial services and healthcare, as examples of what can be done with such a framework.

• Customer Support application covering portable computers and cell phones, using both internal trouble ticked data in SalesForce or ServiceNow, product manuals in PDF or web form and brand related support web sites.
• A medical related demo, applying about ⅓ of the 150 or so rules needed to cover all of the HIPAA laws around confidential medical information.
• A finance demo around investment advisors, insider trading or information leaking
• The design of an upcoming larger demo, supporting questions over 3 years of S&P500 company data (annual reports, proxy reports, websites).

SPEAKER BIO:
Greg Makowski, https://www.linkedin.com/in/GregMakowski has been training and deploying AI since 1992, has been growing Data Science Teams since 2010, has exited 4 startups and has 6 AI patents granted or in various stages. He is currently the Chief of Data Science at Ccube, leading GenAI/AI consulting and framework development for enterprise applications. See also: https://www.ccube.com/genai.

SPONSOR INFORMATION:
From vision to execution, Ccube partners with forward-thinking clients to co-build Apps, Data, and GenAI solutions across industries. Ccube has 10+ service lines, 30+ happy clients, 90% client retention, and saved clients ~50% costs on average.

Ccube has Silicon Valley roots, deep expertise, customer first approach and leverages lean teams for onsite in US and offshore delivery teams in India.

Watch for us also on
https://www.ccube.com/
https://www.linkedin.com/company/ccube-inc/
https://aws.amazon.com/marketplace

As a way to "thank your sponsor", Ccube invites you to share your contact info, and take a brief survey. A summary of the survey results will be shared at the event.

#ACM
#SFbayACM

#GenAI
#GenAIApplications
#EnterpriseApplications
#EnterpriseAIApplications
#EnterpriseGenAIApplications
#RAG
#RetrievalAugmentedGeneration
#Guardrails
#Obervability
#AIsafety
#GenAIcompliance

#Ccube
#LuminAILabs
#IntelliConnect
#Guardiance