AI/ML engineer & CTO · 10+ years · building agentic systems that touch the real world
Medellín, Colombia (UTC−5) · Remote-first, open to relocation
I ship production AI systems — LLM pipelines, multi-agent services, MCP tooling — and lead the distributed teams that run them. Two-time cofounder, currently CTO / Lead AI Engineer at CIMA (Azart Technologies), an AI-driven B2B platform for airlines. Next stop: robotics software — bringing real-time pipeline and protocol experience to ROS 2 and physical systems.
Most of my recent work lives in private company repos. Happy to walk through architecture and code on a call.
- CTO / Lead AI Engineer @ CIMA — LLM-powered document reconciliation for airline financial registries (−60 % manual processing), real-time flight-radar data pipeline (200 ms → 50 ms render latency), MLOps for model CI/CD and observability. Customers include Royal Air Maroc and Centrum Air.
- Building ANSI — a hands-free AI voice tour guide. Flutter client → NestJS gateway → Python AgentScope swarm (6 specialised agents: concierge, planner, researcher, verifier, tour guide, composer) → MCP tool servers (Google Places, Wikipedia, DuckDuckGo, city context) over RabbitMQ; Postgres + Qdrant + MinIO; Kubernetes/Ansible specs for prod.
- Learning ROS 2 — working through a 12-week roadmap (ROS 2 Humble →
ros2_control→nav2). Portfolio repos will land here as they're done.
| Project | What it is | Stack | Status |
|---|---|---|---|
| volve_ddrs | DDR Intelligence — local-first, provenance-preserving extraction, anomaly flagging and cited Q&A over 1,000 Daily Drilling Reports (Volve corpus). Deterministic parser owns facts; LLMs confined to judgement and wording. Runs on one Apple-Silicon machine, zero data egress. | Python, pdfplumber, local LLMs, ADRs | Public |
| ANSI · ansi.world | Multi-agent travel guide platform (see Now). Plus a content-swarm service: Google Sheets MCP + AgentScope swarm that brainstorms, enriches and cross-verifies a content plan under a budget cap. | Python, AgentScope, MCP, NestJS, Flutter, RabbitMQ, Qdrant | Private beta |
| CIMA · cima.systems | Airline operations platform: LLM document reconciliation, real-time flight data, booking-provider and partner integrations, web + mobile + partner APIs. | Python, NestJS, Flutter, LLM pipelines, K8s | In production |
| AI assistant platform (solo founder + AI coding agents) | Metered, memory-bearing AI assistant with tools. Python modular monolith with explicit ports; double-entry token ledger; tests-as-spec; diff-scoped mutation testing in CI; 60+ ADRs; agents implement, I review and merge every PR. | Python 3.12, FastAPI, Pydantic AI, Postgres + pgvector, vLLM | Private, in build |
| Agent team ops | A set of scheduled Claude agent "routines" — PM dispatcher, architect/CTO advisor, code reviewer, QA — that run the day-to-day engineering process for an aviation product: Jira conventions, release cards, digests, scorecards. | Claude agents, skills, Jira, Telegram | Private, running daily |
| Family Compass | Flutter family-tree app: graph model (Person nodes, parentChild/partner edges), four-way swipe navigation, inline add-relative, on-device Drift/SQLite persistence. 134 tests, flutter analyze clean. |
Flutter, Dart, Drift | Local, mock-UI phase |
| MCP servers | Wikipedia and DuckDuckGo search servers deployed on Railway; Playwright browser, Git-notices and personal-finance MCPs for daily use. | Python, mcp SDK, FastAPI |
Mixed |
| fastrtc_speaker_backend | Real-time voice conversation backend: WebRTC in → STT → LLM → Sesame CSM speech out, with TURN and conversation memory. | Python, FastAPI, FastRTC, Hugging Face | Prototype |
Why: I want to build systems that act in the physical world, not just on screens.
What transfers: 100+ Hz data pipelines with sub-50 ms latency, protocol design (REST / GraphQL / WebSocket / MQTT / custom), distributed multi-service architectures, C/C++ and Python, and leading cross-timezone engineering teams.
What I'm doing about it (honest status: actively learning, not yet expert):
- ROS 2 Humble — nodes, topics, services, actions, QoS, tf2, lifecycle nodes
ros2_control— customhardware_interface::SystemInterfacewith a simulated serial/CAN protocolnav2+ SLAM Toolbox — autonomous mobile robot simulation in Gazebo- Planned portfolio: sensor-fusion node (C++,
message_filters, latency metrics) → custom hardware interface → end-to-end AMR simulation
Ideas I'm exploring: an MCP ↔ ROS 2 edge agent so LLMs can call robot and IoT tools through a standard, metered (HTTP 402) interface; multi-agent job/market scouting for the robotics industry; early team-structure work for an exoskeleton project.
Target roles: Robotics Software Engineer · ROS 2 / protocol developer · fleet-management and robot-to-cloud software · autonomous systems.
- invoice_su — Dart client for the invoice.su acquiring API, any platform
- tinkoff_acquiring — Flutter plugin wrapping native iOS/Android Tinkoff SDKs (Apple Pay / Google Pay)
- platinum_list — Flutter SDK for Platinumlist events and ticketing
- yandex_mapkit — Flutter implementation of Yandex MapKit
- weather_kit_native · tabnavigator_plus — smaller Flutter utilities
AI / agents: LLM integration, RAG, agentic systems (AgentScope, Pydantic AI), MCP server & client development, model evaluation, AI document processing, Hugging Face tooling
Languages: Python · Dart · TypeScript · C/C++ · Java · Kotlin
Backend / infra: FastAPI, NestJS, GraphQL, RabbitMQ, Postgres/pgvector, Qdrant, Redis, Docker, Kubernetes, AWS, CI/CD, OpenTelemetry
Mobile: Flutter at scale (6+ years; apps with 100K+ users and 300K+ monthly transactions)
Robotics (learning): ROS 2, DDS, ros2_control, nav2, Gazebo
📜 Track record (click to expand)
| Period | Role | Highlights |
|---|---|---|
| 2025 – now | CTO / Lead AI Engineer, CIMA · Azart Technologies | AI data services for web, mobile and partner APIs; LLM reconciliation −60 % manual time; flight-radar pipeline 200 → 50 ms; MLOps and data-governance frameworks; technical bridge to international airlines |
| 2023 – 2024 | Tech Lead, AI Systems, Shuffle.city · Azart | Led 2 teams / 8 engineers across time zones; 300K+ monthly transactions; CI/CD giving 2× faster releases at 99.5 % reliability; A/B and analytics-driven AI feature validation |
| 2017 – 2024 | Co-founder & CTO, Horum.co | AI-enhanced dev agency: 31+ projects with AI/ML components, 15+ web apps, 10+ automation solutions; feature-flag and A/B frameworks; distributed teams across regions |
| 2020 – 2022 | Co-founder & CPO, Patrick | AI-enhanced productivity app in 3 languages, 100K+ users; 5+ growth hypotheses tested per week; AWS/Kubernetes |
| 2015 – 2018 | Software engineer & product manager, early ventures | Led teams of 3–20 building multi-platform products from idea to market test |
Shipped consumer apps: Shuffle · Flips · Mandarin Markets · Sensera (CBT self-help, retired)
Education: B.Sc. Secure Information Technologies — ITMO University, St. Petersburg (2020) Certifications: AI Agents + MCP (Hugging Face) · Leadership & Management (Udemy) · Project Management, Business-Process Analysis & Modelling (HSE) Languages: Russian (native) · English (advanced) · Spanish (intermediate)
- Hypothesis-driven. Lean-startup practitioner since 2015: 5 failed projects, several that shipped, and the scars to tell them apart early.
- AI agents do the typing; I own the architecture and merge every PR. Tests are the spec, mutation testing keeps them honest, ADRs record the why.
- Distributed by default. I've built and run teams across continents, time zones and stacks — async-first, short calls, written decisions.
- Writing: Task decomposition · Team lead's basics · Leverage collaborative tools · Calls: concise and time-bound
Ask me about: running an engineering team with AI agents · multi-agent systems with AgentScope · building MCP servers · Flutter at scale · real-time data pipelines · lean validation of hardware-adjacent ideas.
I'm looking for: robotics software roles (ROS 2, protocols, fleet management) and cofounders in robotics / physical AI who want a technical partner with production AI and team-building experience.
📫 LinkedIn · coresroman@gmail.com
⚡ Fun fact: 59 countries visited so far — the tour-guide app is partly self-interest. 🌍