AI Agent Infrastructure Engineer
Agent runtimes · Execution evidence · Evaluation
I’m a software engineer focused on AI agent infrastructure and evaluation. I build the systems around AI agents: how they execute tools, manage context and state, preserve evidence, and get evaluated.
Based in Nigeria, I’m the founder of OmniRexflora Labs, Workstream Lead at Flow Research, and an AI Expert Contributor at Snorkel AI. My work connects backend engineering, agent runtime design, and hands-on evaluation of coding agents across languages, toolchains, and execution environments.
OmniCoreAgent — Agent runtime and execution evidence
I build and maintain an open-source Python agent harness that brings tool execution, MCP integrations, memory, context management, workspaces, and background tasks into one runtime.
My current focus is native tool calling, parallel tool execution, streaming, and execution telemetry: connecting user requests, model interactions, tool calls, tool results, context changes, and final responses.
I’m developing the path from that evidence to application-specific offline evaluation—using production behavior to inform controlled tests, rather than treating a recorded trace as an experiment or proof of success.
Workstream — Governed contribution infrastructure
At Flow Research, I lead the engineering of infrastructure for coordinating, verifying, and recording work performed by humans, AI agents, or both.
The work spans identity and authorization, project-scoped permissions, versioned policies, immutable submissions, check evidence, and review/revision workflows. The lifecycle is designed to produce trustworthy contribution records that preserve who did what, under which rules, and with what accepted outcome.
Currently under active v0.1 development. A submission, a passing check, and an accepted contribution are different facts; the system must preserve those distinctions.
I create and review executable evaluation tasks covering terminal-based engineering, long-horizon coding, and research-reproduction workflows. I’ve completed 1,000+ task reviews across projects, with authoring and review contributions spanning Terminal-Bench and other agent-evaluation programs.
This work takes me across languages and toolchains: understanding codebases, reviewing implementations, investigating failures, and checking whether tests establish the behavior required by the task contract.
My work includes task specifications, reproducible environments, reference solutions, executable verifiers, rubric-based assessment, and repeated-run failure analysis. It spans multi-turn pairwise model evaluation—comparing trajectories for correctness, agency, and alignment—and verifier-backed tasks run through Harbor for RLVR-oriented workflows.
A trace is evidence, not an experiment. A verifier is only useful when it checks the right contract.
I care about explicit trust boundaries, reproducible tests, meaningful failure analysis, and preserving the evidence needed to explain what happened.
My standard for engineering ownership is straightforward: explain the architecture, justify the tradeoffs, and reason about failure modes—not just produce working code.
Languages I’ve worked with: C, C++, C#, Rust, Go, TypeScript, Python, Bash, Perl.
Backend and runtime: FastAPI, asyncio, AnyIO, PostgreSQL, Redis.
Infrastructure and integration: Docker, Linux, MCP, REST APIs, object storage.
Evaluation: Harbor, executable verifiers, rubric design, trajectory analysis, reproducibility, and failure analysis.