Agents, observation, and the Agentic Loop next steps

AI Orchestration: From Custom Skills to Autonomous Hedge Fund Operations

Most asset managers treat AI as just a chatbot, failing to bridge the gap between an LLM's general reasoning and the specific, high-stakes workflows of their actual day-to-day.

In this episode of The Hedgineer Podcast, Michael Watson sits down with Jhanvi Virani, COO of Hedgineer, to discuss the practical mechanics of deploying AI within hedge funds and asset managers. Jhanvi details her experience shadowing a CIO to translate their cognitive investment process into a digital skill—a structured framework that allows Claude to synthesize fragmented data from order management systems, SharePoint research, and consensus estimates into polished, institutional-grade outputs in a one-day turnaround. We move beyond simple prompting to explore the "Agentic Loop," discussing how local schedulers and the Claude Agent SDK are enabling systems to run autonomously 24/7.

The conversation also covers the technical nuances of the Claude Ecosystem, comparing developer-centric Claude Code with user-friendly Claude Cowork. Jhanvi shares her on-the-ground findings regarding the limitations of local vs. remote execution and why building a secure, server-side environment is the ultimate bottleneck for scaling AI intelligence across a firm.

Key Takeaways

• The Skill-Based Unlock: How shadowing investment professionals allows engineers to map complex and manual research workflows into automated skills that produce consistent, high-polish one-pagers.
• Claude Code vs. Cowork: A breakdown of why developers prefer terminal-based workflows for multitasking, while non-technical users leverage Cowork for scheduled tasks and visual connector management.
• Building "AI Native" Infrastructure: The 0-to-1 process of auditing fund workflows, building custom MCP (Model Context Protocol) connectors for legacy data vendors, and establishing organizational agent management frameworks.
• The Self-Healing Feedback Loop: Using usage analytics and "meta-agents" to observe behavior, evaluate performance, and automatically suggest system improvements, creating a self-sufficient AI framework.

Timestamps

00:00 - Introduction and the role of skills in unlocking automation 

04:15 - Evolving daily workflows with Claude Code and Cowork 

08:42 - UI vs. Terminal: Optimizing screen real estate and parallel sessions 

14:30 - Testing the bounds: Automating expense reports and attachment limitations 

17:45 - Windows vs. Linux runtimes and the "Local Scheduler" in Cowork 

22:10 - The Agentic Loop: From Claude Agent SDK to OpenClaw deployments 

29:40 - CIO Shadowing: Translating a day of research into a custom AI skill 

36:50 - The future of autonomous analytics and observation agents 

43:15 - Deliverables for becoming AI Native: Audits, MCP servers, and data warehouses 

51:00 - AI Personification: Authenticity in communication and the risk of "AI slop." 

64:20 - Team expansion in Bangalore and the tech-focus of South India

Guest Bio: Jhanvi Virani is the COO of Hedgineer, where she oversees the deployment of AI infrastructure and automation for institutional asset managers. She specializes in bridging the gap between technical LLM capabilities and high-level investment workflows.

Host Bio: Michael Watson is the founder of Hedgineer and host of the podcast, focusing on the intersection of data science, AI, and hedge fund technology.

Links & Subscribe

Subscribe for weekly analysis on AI and Asset Management.

youtube.com/@hedgineer

Hedgineer.io

Hosted on Acast. See acast.com/privacy for more information.