Current Work · Nov 2025 — Present
Evaratus
Founding Member of Technical Staff (Intern)
Building the reinforcement-learning environments, tool-use infrastructure and graders used to train and evaluate LLM-based AI agents.
- 30+
- Multi-agent RL environments
- 1 TB+
- Structured data behind the tool-use API
- 5
- Enterprise data sources
- Teams
- Led through planning, design & reviews
A What I built
- Built 30+ multi-agent reinforcement-learning environments in Python, including multi-app environments where agents complete workflows spanning several enterprise tools — used to train and evaluate LLM-based AI agents.
- Engineered a tool-use API over 1 TB+ of structured data from 5 enterprise sources, letting agents query state and execute long-horizon actions inside reproducible Docker containers.
- Designed task suites, reward functions and automated verifiers, pairing LLM-as-judge grading with programmatic checks to score agent trajectories against a standardized benchmark.
- Led multiple engineering teams through task planning, technical design and code reviews, mentoring teammates on simulation and grading design.
- Python
- Docker
- LLM-as-judge
- Reinforcement learning
- Multi-agent orchestration
- Data pipelines
B Architecture
source5 enterprise sources
Structured data from five enterprise tools — the raw state agents will work against.
C Leadership & behaviour STAR
Leading engineers through the build
- Situation
- Building 30+ environments meant several engineers working in parallel on simulations, tasks and graders.
- Task
- Keep the work planned, the designs coherent and the quality consistent across teams.
- Action
- Led task planning and technical design, ran code reviews, and mentored teammates on simulation and grading design.
- Result
- Multiple teams shipped environments against a shared design and grading approach.
Grading agents you can trust
- Situation
- Agent trajectories are long and open-ended — hard to score with rules alone.
- Task
- Make evaluation both nuanced and reliable.
- Action
- Designed task suites, reward functions and verifiers that pair LLM-as-judge grading with programmatic checks.
- Result
- Agent trajectories scored against a standardized benchmark.