AI Agent Architect

Building an agent is easy. Getting it to produce the same right answer three times in a row — on real enterprise data, in a regulated environment, without a human checking every output — that's the actual problem. You've shipped AI agent systems to production on real, unclean data. Not a demo dataset. You know the accuracy cliff. You know why prompting cannot fix semantic problems. You've built systems that don't rely on the model getting it right every time, and you've built the judgment to know when to ship anyway. You're not looking for a well-defined architecture to implement. You're looking for the unsolved problem — and the mandate to build the solution that becomes the standard.

edisyl builds AI solutions that turn messy institutional data into decisions, workflows, and outcomes. We came out of blockchain data infrastructure — 8 years, 20+ chains, 700M+ resolved wallets — and now deploy that capability to enterprises navigating the same challenge: how to make their data work for them at scale, without armies of analysts. We have active deployments with Fidelity and Interlochen, a proven architecture, and inbound from firms that need what we've built. The technology works. What we're building now is the enterprise motion around it.

• •Design and build the architecture for AI agent workflows — planning loops, tool use, memory, retrieval, and human-in-the-loop checkpoints
• •Evaluate, integrate, and fine-tune foundation models and LLM APIs for specific enterprise use cases and data types
• •Define standards for agent reliability, observability, and failure modes in production deployments
• •Collaborate with Forward-Deployed Engineers to translate what's working in client environments into reusable platform components
• •Build internal tooling and eval harnesses to assess agent quality, hallucination rates, and task completion
• •Make principled, documented architectural decisions — and stay current enough with the ecosystem to know what to adopt and what to ignore

• •Competitive base salary and meaningful early-stage equity. This is a foundational technical role and we price it that way. We'll be transparent about the full picture in our first conversation.

• •6–10 years building production AI or data systems — not prototypes; systems that run reliably at scale under real conditions
• •Deep hands-on experience with multi-agent architectures: context windows, memory management, dependency graphs, and where things break in practice
• •Strong Python and familiarity with agent frameworks — LangChain, LlamaIndex, AutoGen, or equivalent — or a clear, documented opinion on why you built your own
• •Practical experience with RAG architectures, vector databases, and context window management in production settings
• •Experience deploying LLM-powered systems in enterprise contexts — data security, access controls, audit logging

• •LLM failure mode literacy. You know the accuracy cliff. You know why prompting cannot fix semantic problems. You build systems that don't rely on the model getting it right every time.
• •Production instincts. You don't consider something done until it's been wrong three times and you've fixed it twice — and you've built the judgment to know when to ship anyway.
• •Strong opinions on agent design. You have a clear answer to why most agent architectures fail at enterprise scale — and you've built something that doesn't.
• •Systems thinking. You design for failure modes first. Happy paths are not the interesting problem.