The Video Signal

AI Engineer

AI Engineer NYC Livestream | Bridgewater, Anthropic, Two Sigma, Arize, Coatue, Nous, Gen. Intuition

Published 2026-10-13 · Duration 8:07:08

Summary

This technical summary covers advanced architectural patterns for building, governing, and scaling production AI systems, primarily within the financial services sector. Key themes include establishing robust data foundations (e.g., The Vault), implementing multi-layered agent frameworks, and adopting rigorous governance controls (Agent Registry, Sandbox, Eval Layer) to manage non-deterministic AI behavior. Speakers emphasize that the modern AI engineer's role shifts from writing features to building the entire 'Build-Eval-Ship-Monitor-Improve' loop, making observability and continuous evaluation paramount for enterprise adoption.

Download summary

Key takeaways

  1. The Shift to Agentic Systems 23:40

    The focus is moving from reactive tools to proactive agents that can act as participants in complex processes. This requires building foundational layers that allow agents to operate across multiple systems and data sources.

  2. Data Foundation is the Core Asset 10:10

    For agents to be impactful, they require a centralized, governed data repository (like Bridgewater's Vault) that indexes and aligns both internal and external unstructured data to an internal ontology, ensuring context and time-period specificity.

  3. Scaling Agents Requires a Five-Layer Framework 31:20

    To productionize agents, a structured approach is needed: 1) Connectors (access to systems/data), 2) Skills (business rules/processes), 3) Identity/Permissions (who the agent acts as), 4) Approval/Sandbox (human-in-the-loop), and 5) Eval Layer (regression testing).

  4. Governance and Risk Management are Non-Negotiable 37:20

    In regulated industries, controls must balance safety and innovation. Best practices include implementing an Agent Registry, using sandboxes, and employing deterministic guardrails (hooks) to prevent overreach and maintain auditability.

  5. The AI Engineer's Role: Building the Loop 41:40

    The role is evolving to building the entire 'Build-Eval-Ship-Monitor-Improve' loop. This involves setting up systems for agents to consume observability data (telemetry) and using specialized tools (like Arise's Signal) to automate issue detection and continuous improvement.

Technical details

  • Vector Database Filtering Techniques 1150s

    When designing search for agents, advanced filtering is critical. Techniques include: 1) Post-filtering (vector index first, then metadata filter - fast, but lower recall); 2) Pre-filtering (filter first, then calculate distances - high recall, but slow); and 3) TurboPuffer's native filtering (combines speed of post-filtering with high recall of pre-filtering).

  • Foundation Model Architecture for Tabular Data 2750s

    For financial data (e.g., transactions), the model can be pre-trained on structured/tabular data (transactions) rather than just language tokens. The goal is to learn a generalized representation of customer behavior across multiple dimensions (e.g., debit, credit, check) which can then be fine-tuned for downstream tasks like fraud detection or credit scoring.

  • Agent Development Lifecycle 1880s

    The process involves moving from initial experimentation to production and scale. Key components include: Connectors (data access), Skills (business logic/workflows), Identity/Permissions (role definition), Sandbox/Approval (human oversight), and Eval Layer (regression testing).

  • Software Quality and Disposability 2000s

    Code quality is fundamentally about improving 'understanding' and 'execution' of system behavior. The concept of 'disposable software' suggests that if the consequences of failure, cost of replacement, and dependencies are minimal, traditional engineering rigor can be reduced or eliminated.

Mentioned resources

  • Arise (Presenting Sponsor/Platform)
  • Bridgewater (Financial Institution/Data Platform)
  • The Vault (Data Platform)
  • Anthropic (AI Model Provider)
  • Claude (LLM/Product)
  • Haiku, Sonnet, Opus, Fable (LLM Model Family)
  • Two Sigma (Hedge Fund/AI Architecture)
  • Signal (Observability Agent)
  • TurboPuffer (Vector Database)

Channel & topics

Watch on YouTube · Back to latest

This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.