Expertise

AI Systems Engineering

Retrieval pipelines, evaluation harnesses, and agentic tools inside real products.

My interest in AI is less about training models and more about making them dependable inside a product that people rely on. At Easy8 I worked in the R&D team behind the AI assistant for a corporate task-management platform, where my remit ran from model deployment and inference through the full RAG pipeline to the Python service layer that connected it back to a Ruby on Rails system. The part I care about most is evaluation: without datasets, judges, and tracing, "the assistant feels better now" is not an engineering claim. Alongside that, I use my own projects to explore ideas that would never fit an enterprise roadmap.

Evidence

What that looked like in practice

Model deployment & inference

Deployed and compared self-hosted LLMs on AWS and CUDA hardware across Ollama, vLLM, llama.cpp, and TGI, selecting an inference stack per workload instead of defaulting to a hosted API.

RAG pipeline, end to end

Built the retrieval layer over a corporate knowledge base: embedding-model serving (TEI, vLLM, SentenceTransformers, BGE, E5, Jina, OpenAI), crawlers and spiders to aggregate scattered internal content, vector storage on Chroma, Pinecone, and pgvector, and the normalisation, chunking, and overlap tuning that decides whether retrieval is genuinely useful.

Quality measurement

Built the tooling that showed whether a change actually helped: curated datasets, LLM-as-judge and code-based evaluators, and tracing via LangSmith and Promptfoo — so assistant tools were compared on evidence rather than impressions.

Agentic tooling & integration

Designed AI assistant tools and their API (Python, FastAPI, LangChain) for integration into the Rails platform, including agents that turn a natural-language request into a safe, constrained query or a generated report.

Personal exploration

In StillMe, a multi-provider LLM proxy and prompt engineering (validated with Promptfoo) model a conversation with your own self five years from now, grounded in analysis of the user's writing style. In Awaken Protocol I prototyped an architecture for LLM-generated automation, letting a player describe fleet behaviour in natural language or by voice and have the tooling built for them on the fly.

Tools & practices

Python FastAPI LangChain vLLM / Ollama / TGI / llama.cpp RAG TEI SentenceTransformers pgvector Chroma Pinecone LangSmith Promptfoo AWS CUDA

Where this came from

Easy8 Software (2020–2024)
Awaken Protocol (2024–present)

Full work history

Contact

Want the detail behind any of this?

Happy to talk through the specifics — what worked, what didn't, and what I'd do differently.

panov.eduard.k@gmail.com