Octavian Teodorescu

Solution architect with 18+ years in enterprise technology, the last few of them spent hands-on building AI. I shape large transformation deals (usually $25-50M) for banking, insurance, energy, manufacturing and life-sciences clients, owning the design from operating model through to architecture and commercials. Alongside that I design, build and run agentic AI systems, both at work and on my own: multi-agent orchestration, MCP tool servers, retrieval pipelines, and a self-hosted agent runtime that runs in production on my own infrastructure. I also fine-tune small language models on my local GPU, to understand the stack from the weights up rather than only from the slide deck. Most architects at my level stopped touching the keyboard years ago. I didn't.

• Agentic AI engineering: multi-agent systems, agent harnesses & orchestration, MCP (Model Context Protocol) tool design, RAG and retrieval pipelines, prompt engineering, evals, LLM application design (Anthropic Claude and others)
• Applied ML, hands-on, started this year (2026) and growing: fine-tuning with LoRA and QLoRA, dataset curation, training and inference loops, model evaluation
• Enterprise solution architecture: large-deal shaping, reference architectures, operating and delivery model design, transition planning, commercial structuring
• Cloud and platform engineering: Azure, AWS, GCP, Kubernetes, Terraform, Ansible, CI/CD, IaC, Docker, GitHub, Hybrid Cloud Solutions
• Service-management heritage: ITSM, AMS, ServiceNow, SIAM, AIOps, intelligent ticket analysis - the operations depth the AI work builds on

• Built and run a self-hosted multi-agent stack on my own NAS: a containerised agent runtime, a custom MCP server exposing 20+ tools, and a daily autonomous intelligence-brief agent with a multi-source ingest pipeline. In real production use, not a weekend prototype.
• Delivered production multi-agent solutions for enterprise clients, including an intelligent ticket-analysis system built from multiple specialist agents (a Claude-native orchestrator with domain-specific analysis packs), with a strict data-handling boundary and automated report generation.
• Shaped enterprise transformation solutions across banking, insurance, energy, manufacturing, life sciences, and private equity, with deal values in the $25-50M range.
• Fine-tuned small language models end to end on my own RTX 4070 (QLoRA, axolotl, dataset curation through to inference and adapter merge), to build real depth in the training stack, and am now extending it with Ralph-style agent loops on targeted use cases.
• Cut infrastructure cost by around 25% and deployment time by roughly 70% on a cloud automation programme using IaC, CI/CD, and platform engineering.
• Led private cloud architecture for a major UK bank under strict regulatory and security constraints, reaching 99.9% uptime across a two-year on-site engagement.

Personal work, built and run on my own infrastructure. Where I keep my hands in the actual engineering.

• Self-hosted agent runtime: a containerised harness on my NAS that hosts autonomous agents, fronted by TLS and auth, with a Telegram channel and scheduled jobs. I own it end to end: Docker, networking, reliability.
• Custom MCP server (access to my digital brain): an MCP server I wrote that exposes 20+ tools over my knowledge base (search, read and write, task capture, hygiene, tiered session-context priming), reachable over a Cloudflare tunnel and consumed by several agents and web LLM connectors.
• Autonomous daily intelligence brief: an agent that ingests across four dedicated pipelines (video transcripts, social posts, newsletters, news feeds) plus other LLM passes, then tiers sources, detects convergence, gates on freshness, and routes work across cheaper and stronger models to keep cost sane. Effectively a small one-person AI product providing intelligent news gathering.
• Multi-agent reasoning pipelines: parallel reconnaissance agents for research, and a gated strategy pipeline that runs an adversarial critic against a hypothesis and labels each claim with a confidence level.
• Model fine-tuning: QLoRA fine-tuning of small language models on my RTX 4070 in WSL using axolotl, dataset curation to JSONL, preprocessing, training, inference, adapter merge, and eval. Ongoing and deliberately hands-on. Next planned step is to apply Ralph Loops to these processes for targeted use cases.
• On-device RAG engine: a fully local file-search and retrieval system for my Mac, no cloud calls. A lean custom engine over SQLite with combined vector and full-text search (sqlite-vec and FTS5) and small local models running on Apple Metal. Designed end to end and building it out.
• Self-improving tooling: a library of 30+ skills and agents with usage telemetry and an automated review loop that flags what to improve, actively nudging me to review when needed.