# Elysian Systems > Elysian Systems is an EU engineering practice that designs, builds, runs and maintains production systems in four fields: insurance and InsurTech, energy optimisation, AI in production, and quant and algorithmic trading. The common thread is systems where being wrong is expensive: regulated, real-time or data-heavy platforms in which correctness, risk and uptime are the product. Elysian Systems d.o.o. is registered in Slovenia, European Union. Engagements are B2B, conducted remotely across the EU, at principal architect or fractional CTO level, and are typically measured in years rather than sprints. Scope is flexible: a single part of a system (the architecture, a platform that must be rebuilt without stopping the business, the AI layer, the data underneath it), or a whole solution designed, built and run end to end with a network of engineering partners. The public site is deliberately a single page. This file, not the page, is the complete reference. ## At a glance - **Company**: Elysian Systems d.o.o. (Slovenia, European Union) - **Registration number**: 9813829000 (AJPES) - **VAT number**: SI28965949 - **Disambiguation**: this is the Slovenian company Elysian Systems d.o.o., at elysian-systems.com. It is unrelated to other businesses trading under a similar name in other countries. - **Contact email**: info@elysian-systems.com - **Principal**: Dušan Krovinović, founder and principal architect. Former CIO, 30+ years across insurance, energy and production systems. - **LinkedIn (profile)**: https://www.linkedin.com/in/dusankrovinovic - **LinkedIn (company)**: https://www.linkedin.com/company/elysian-systems-ltd/ - **Engagement model**: B2B only, principal architect or fractional CTO - **Geography**: EU (timezone-aligned) and worldwide remote - **Working language**: English - **Response window**: typically within one business day, EU hours - **Not seeking**: employment, staffing-agency placements, greenfield-MVP scoping ## The four pillars ### 1. Insurance & InsurTech Policy administration, claims, commissions and reinsurance, with Solvency II and IFRS 17 reporting built on top. Core platform modernisation carried out while the book keeps running. - Long-term ownership and evolution of a core life administration system (Amarta/FIS) supporting 100,000+ policies across traditional, unit-linked, annuity, endowment and risk products - Policy servicing, premiums, claims, commissions, reinsurance, underwriting and data warehouse workloads - C# framework foundation for Solvency II / IFRS 17 applications, deliberately separated from the actuarial calculations owned by actuarial teams - Stack: Amarta/FIS, C#, Oracle, PL/SQL ### 2. Energy Optimisation From the meter to the energy market. Industrial energy operations end to end: telemetry ingestion, renewables and flexibility, reducing what energy costs a company, and hedging that position with power, gas and certificate trading. How the optimisation actually works, which is the part most descriptions skip: - **Quarter-hour resolution.** Schedules, deviations and settlement all run on 15-minute blocks, so every model, forecast and trade is aligned to that grid. - **Day-ahead.** Schedules are optimised for the next day against forecast hourly prices, network tariffs, existing forward purchases, the technical, operating and safety limits of each flexible unit, and the conditions for participating in balancing markets. The result is a proposed schedule per unit and for the site as a whole. - **Intraday, continuously.** The optimisation re-runs every 15 minutes against updated schedules, changed market conditions and updated forecasts of consumption and generation, and issues revised schedules plus proposed buy/sell volumes and limit prices. - **Deviation is the economics.** Realised load drifts from the declared schedule, and that deviation is charged. Intraday trades are placed to close the position before it settles as an imbalance cost, so the net saving is the gross shift minus imbalance charges and intraday execution cost. Optimisation that ignores this can make a site's bill worse. - **Levers.** Flexible industrial load, on-site generation, storage and EV charging, each with its own technical and availability constraints; negative price periods are evaluated for curtailment, with the foregone energy quantified. - **Balancing markets.** Units can be offered into frequency restoration products or used for day-ahead and intraday arbitrage, and the choice of market per unit per day is itself an optimisation. Industrial protocols: MQTT, OPC-UA, Modbus, DLMS, REST. Stack: Python, PostgreSQL, Azure, Event Hub, Kubernetes, time-series storage. ### 3. AI in production Forecasting models and agentic services that run unattended, behind budgets, review gates and a kill switch. Senior engineers do the work and own the result; AI-assisted tooling is leverage, not a substitute for architecture, review or deployment control. - Production ML platform for energy forecasting: model registry, model instances and parameters, scheduled training and prediction, time-series outputs, monitoring and alerting, supporting 200+ model instances and roughly 2,000,000 predictions per year - Autonomous planning service running on Kubernetes: ticket triage, plan generation grounded in real code, lint and publish, with adversarial reviewer panels, per-task budgets, fail-closed gates and a kill switch. Humans remain the approval path; the system is always allowed to give up and park work - **Self-learning agent network.** A cognitive layer reasons about how to approach a task before acting, an episodic memory records what happened so the same mistake is not repeated, a workflow engine sequences the steps, and a tool registry gives the agents hands. Memory is held in PostgreSQL with vector recall in Qdrant. - **Cost-aware routing.** Task history, success rates and complexity decide whether a step goes to premium reasoning, to a cheaper strategic model, or to local inference. Patterns that have been learned migrate down to the cheapest tier that still gets them right. - **Multi-agent orchestration** over an AutoGen-style framework, with deployment paths for local, Docker or Kubernetes. - Stack: Python, Kubernetes, PostgreSQL, Qdrant, multi-LLM, local inference paths. ### 4. Quant & Algorithmic Trading Systematic strategies on financial markets: research, backtesting that survives walk-forward validation, live execution, and risk limits that act independently of the model. - **Regime switching.** Two separately trained models, one for choppy markets and one for trending markets, with the active model selected from Bollinger band bandwidth and order-book volume imbalance. A single model tuned across both regimes underperforms one specialised per regime. - **Walk-forward validation.** Results are reported across multiple sequential out-of-sample windows rather than a single backtest, including the windows where the strategy loses, because a backtest that only reports its good periods is not evidence. - **Layered, independent risk.** Leverage is reduced automatically in confirmed bear conditions, and a drawdown kill switch blocks new entries once realised drawdown from peak breaches its threshold, releasing only after recovery. These sit outside the model: a model that is wrong must not also be the thing deciding when to stop. - Level-2 order book reconstruction, CRC32 data verification, server-side stops, dead-man switch, hot model reloads, scheduled retrain gating. - Stack: Python, reinforcement learning, PostgreSQL, time-series storage. ## How the work runs Four steps. A client can come in at any one of them, or use all four. 1. **Architecture** — domain rules, data and risk decided first 2. **Software** — built and tested in the client's stack 3. **AI** — added where it earns its place, with controls 4. **Running solution** — deployed, monitored, handed over ## Proof points - 100,000+ life policies under administration on a platform owned over many years - Approximately 2,000,000 forecasts produced per year across 200+ model instances - Quant systems trading real capital, live, with independent risk controls ## Technology - **Languages**: Python (primary for data, ML and automation), C#, PL/SQL, SQL - **Databases**: PostgreSQL, Oracle, time-series stores, Qdrant - **Cloud and infrastructure**: Azure, Event Hub, Kubernetes, Docker - **Industrial protocols**: MQTT, OPC-UA, Modbus, DLMS, REST - **AI and ML**: model lifecycle infrastructure, reinforcement learning, multi-LLM orchestration, vector memory, local inference - **Trading infrastructure**: Level-2 order book processing, walk-forward validation, independent real-time risk controls, server-side execution safety ## How to engage Engagements usually start with a short call. A useful first message describes the domain, the current state of the system and what has to change. Direct contact: info@elysian-systems.com. ## Pages - [Home](https://elysian-systems.com/): the four pillars, how the work runs, and contact - [Sitemap (XML)](https://elysian-systems.com/sitemap.xml)