AI tools for Forward Deployed Engineering — Vasuman Moza, Varick Agents

The speakers emphasize the use of dependency graphs to represent enterprise workflows, noting that while various graph database solutions exist (mentioning Postgres and several vendors), the specific technology is less important than the concept. They explain that process owners prefer a dependency-driven approach where approvals are sequential, preventing individuals from proceeding without necessary sign-offs. This structured representation facilitates AI automation.

The discussion highlights a surprising limitation of frontier models like Claude: they often produce verbose and unfocused outputs. To address this, Varick agents post-train their own models on top of open-source foundations (specifically mentioning Kimiko 26). This custom training aims to achieve a balance between detailed analysis and clarity, mirroring the ability of experienced consultants to prioritize client concerns.

To effectively utilize knowledge graphs, Varick agents create a reinforcement learning (RL) environment. This environment is used to train custom tools designed to navigate the graph and extract relevant context. These tools address specific challenges like resolving ambiguity around names (e.g., multiple 'Mikes' in a company) and identifying redundancies or violations within the workflow.

Varick believes that the biggest bottleneck in AI implementation isn't technology, but a lack of deep understanding of how businesses operate. They address this by employing forward deployed engineers and strategists who conduct audits within client companies to learn their processes from the inside out. This foundational knowledge informs subsequent agent building and ensures solutions are tailored to specific business needs.

Varick argues the primary bottleneck isn't AI execution anymore, but rather a company’s ability to understand its own business processes. He highlights that every business and consumer operates differently; for example, sales departments in healthcare versus SaaS companies function uniquely. Extracting this context from employees and feeding it into AI models remains challenging due to the difficulty of translating nuanced knowledge into API calls.

Forward-deployed engineers are crucial for bridging the gap between current operations and future AI-driven processes. Their role involves understanding existing workflows, identifying pain points, and re-envisioning how AI can transform them. This goes beyond simply applying AI tools; it requires fundamentally changing operational processes to leverage AI effectively.

The initial step for forward-deployed engineers is meticulously mapping how humans currently perform their work. This involves embedding engineers within specific departments, such as finance, and interviewing process leads to understand both standard procedures and the often undocumented ways things actually get done when issues arise. For instance, an FDE might discover that Sarah in AP handles a workflow one way but relies on Chris for reconciliation, adding significant cycle time.

Varick cautions against simply applying AI to existing, flawed processes. He points out that this is a key reason for the lack of widespread ROI in AI adoption. Instead, he emphasizes the need to re-engineer processes *around* AI, which requires understanding the underlying issues and designing workflows that leverage AI's strengths effectively.