ankush wadehra

I’m an applied AI engineering candidate focused on building agentic workflows that are measurable, reproducible, and production-reliable. I’m especially interested in LLM tool calling, structured outputs, and RAG concepts that translate data into dependable next actions.

I built Agenta, a LangGraph/FastAPI autonomous agent platform with structured routing, streamed execution, a 4-tool registry, and loop guards. I aimed for clear validation paths—so agent behavior can be controlled, traced, and improved rather than treated as a black box.

On the evaluation side, I’ve produced and reviewed 4000+ verifiable AI evaluation tasks for coding agents, including terminal debugging, automated tests, trace review, and rubric-based assessment. I also reduced review cycles by 70% through clearer prompts, stronger tests, and documented failure cases for agent and coding workflows.

Alongside evaluation, I architected a stateless agent orchestration npm package for plug-and-play multi-agent coordination, with concurrent execution support that cut inter-agent communication overhead by ~40%. I also contribute to computer vision and deep learning prototypes—engineering PyTorch/OpenCV pipelines achieving over 90% detection accuracy and reducing iteration cycles by about 35%.