I'm Antonie Chirilus, an R&D engineer at Keysight Technologies. I build the parts of LLM systems that have to be correct — guided generation, schemas, validators, self-healing loops, local inference. The unglamorous half of agentic AI. I write about it here.
Role
R&D · AI Engineer
Focus
LLM reliability, agents
Based
Bucharest, RO
Stack
Python · llama.cpp · MCP
01 / 04 · Writing
Field notes from production LLM work — what breaks, what holds, and the math that decides which.
I spent a week chasing latency on a local LLM setup. The bottleneck wasn't the GPU, the model, or the quantization — it was the prefill stage doing the same 40K tokens of work on every request.
Why I'm starting this, what to expect, and the kind of writing I want to do here.
02 / 04 · Selected work
Persistent memory for LLM agents, modeled on human cognition. Seven typed stores — episodic, semantic, procedural, entity, working, summary, buffer — each with its own update strategy. 44 tests, live Streamlit demo.
Drop-in MCP middleware that catches silent tool-output drift — when Slack reorders threads, an API renames a field, a DB schema shifts — before it cascades through your agent chain. SBERT embeddings + PSI on the projected distribution.
A five-agent pipeline that turns a one-line requirement into a working CrewAI repo on GitHub. Architect, codegen, test-writer, reviewer with self-correction, deployer.
An autonomous code-audit swarm. Navigator + Analyst agents on AutoGen 0.4, tool access through MCP, walking a remote repository and producing a structured improvements report.
03 / 04 · Experience
R&D Engineer · AI
Keysight Technologies · Bucharest
Correct-by-construction generation on the Visibility Orchestrator. Outlines, Pydantic, llama.cpp, RAG grounded in network data.
AI Engineer · Intern
Keysight Technologies
Production RAG over Keysight's audit corpus. Function-calling agents on Azure OpenAI, retrieval through Azure Cognitive Search, FastAPI.
ML Engineer · Contract
Roglia SRL · Remote
Custom AI chatbot for public institutions. CrewAI, LangChain, Voiceflow, Groq + OpenAI inference, Grafana for observability.
B.Sc Computer Science
University of Bucharest
Finishing this year. Prior: several years of mathematical olympiad at the national stage.
04 / 04 · Contact
If a post here was useful, wrong, or worth arguing about — email me. I'm also open to AI engineering roles, full-time or contract, hybrid in Bucharest or remote on European hours. The shorter the email, the faster the reply.