De Jesus’s personal project captures prompts, responses and tool calls in Langfuse. A hook processes new messages after each assistant response and groups traces by session.
A pipeline described by De Jesus that evaluates models and whole agents against financial datasets, producing pass-or-fail results and reviewable traces. Human approval remains separate.
An official open-source collection introduced by De Jesus and Al Brown. Its initial 28 prioritized rules covered schema design, query optimization and ingestion.
Doneyli De Jesus is a solution architect at Langfuse, where he works on AI and data infrastructure. His projects help engineers inspect coding agents, evaluate AI applications, and make deployment decisions with evidence they can review.
From data infrastructure to AI engineering
De Jesus began in IT help desk and has spent more than two decades working with data and AI systems, including at Deutsche Bank, Verizon, Cloudera, Elastic, and Snowflake. At Snowflake, he became a principal and executive advisor and helped lead the company’s first AI go-to-market effort. He joined during the pandemic, when client engagement and dealmaking had to move online. His farewell note also describes mentoring colleagues and working on cross-functional programs.
At Langfuse, now part of ClickHouse, he works with an open-source platform for tracing, evaluating, and managing prompts in LLM applications. His GitHub profile describes how he learns by building with tools and sharing what he discovers about using them in production.
Making agents easier to inspect and evaluate
Coding-agent observability. De Jesus created a personal Claude Code and Langfuse template to capture prompts, responses, and tool calls. After each assistant response, a hook reads new conversation messages and sends traces to a self-hosted or cloud Langfuse instance, grouping turns by session. Developers can then inspect the agent’s tool inputs and outputs alongside the conversation.
Deployment gates for financial AI. In work on LLM and agent evaluation, De Jesus describes a pipeline that tests financial applications against datasets, scores their outputs, and produces a pass-or-fail result with traces for review. It evaluates whole agents as well as individual models: a correct numerical answer, for example, can still fail the tool-use check if the agent should have used a calculator. Reviewed production failures can become new test cases, while prompt changes can trigger another gate run. The pipeline supplies evidence for human decisions; it does not issue a certificate or replace independent approval.
Database expertise for coding agents. De Jesus and Al Brown introduced ClickHouse Agent Skills, an open-source collection of instructions covering schema design, query performance, and data ingestion. The initial release packaged 28 prioritized rules so an AI assistant could draw on specialized guidance when needed. A query or schema may work in a small test yet cause trouble at scale if it makes poor choices about ordering, joins, or insert batching.
Annabell Schäfer and Doneyli De Jesus use a banking assistant to show how traces, expert review, evaluators, and deployment gates can reveal failures that ordinary system metrics miss.
Trace tool calls and intermediate steps: an agent can report a successful action that it never performed.
▶ Watch ↗