Dávid Natingga, Ph.D.

I implemented a simple quantum simulator in Rust with a quantum complete set of gates. This simple simulation works by computing the complex probability amplitudes for all the possible qubit measurements. The code deals with well-known concepts in quantum mechanics, such as the Bloch sphere, superposition, entanglement, the uncertainty principle, collapse, and more.

I designed algorithms and models to disaggregate the active power consumption data from household and commercial sites, at a 15-minute resolution, into the consumption of individual appliances, including fridges, dishwashers, washing machines, heating, air conditioning, electric vehicle charging, and others.
The challenge was that the labeled data was limited and only at a 15-minute resolution. Therefore, an approach utilizing representational learning on large amounts of unsupervised data was employed.

I researched and created algorithms and models to perform non-intrusive load monitoring in low-voltage networks. The data was at a 1-minute resolution and included the power quality metrics such as active power, reactive power, and harmonics. The objective was to detect the presence, count, and power consumption of various distributed energy resources (DERs) such as electric vehicles, heat pumps, and solar panels. The labels across the full power quality spectrum were limited. So, simulation of the remaining metrics based on the active and/or reactive power was used to improve the results.

I turned a meter into a smart meter using its raw current and voltage data to track devices for consumption and predictive maintenance.

I challenged state-of-the-art research and invented new robust algorithms and techniques to reach near-perfect performance for device disaggregation (accuracy over 99%). The project required working within a custom-made meter's hardware limitations and specifications. This involved the calibration and robust measures of the power quality metrics.

I designed an algorithm to infer the network topology of a low-voltage network from the real-time time series power quality data from the high-frequency meter at the transformer level. The algorithm took advantage of diverse device properties, such as fixed-resistance and power devices. The method was based on the simulation and modeling of electrical engineering principles.

These were my tasks for this project:

• Created data validation algorithms to validate imported data and to detect and correct corrupted time-series data.
• Built an automatic report generation tool to provide an overview of the quality of the data, pinpoint specific problems in the data, and detect problems with the devices captured by the data.
• Designed a scalable SQL database structure based on the TimescaleDB extension to run research experiments at scale.
• Found bugs in the client's infrastructure and advised its simplification and decoupling to make it more robust and more efficient for working between development and data science teams.
• Created novel algorithms to detect mechanical misconfiguration of the sensor devices and correct the data from the misconfigured devices.
• Conducted an extensive statistical study on the uncertainty and confidence intervals of the data received from the monitoring devices.
• Developed a strategic roadmap and new architecture for advanced analytics.

Frequency Containment Reserve is an indispensable tool for maintaining stability in an electrical power system.

I made a Frequenz SDK app to perform FCR using a microgrid battery. I probabilistically computed the configuration that minimizes the energy market risks, maintains high microgrid stability, and preserves the battery. The project also involved generating the grid frequency for the simulation and understanding the patterns and distribution of the grid frequency.

Using linear optimization (Google OR-tools), found an optimal battery size based on the battery price, charge rate, capacity, and the opportunities with which the battery can be monetized, including energy trading, balancing, and storage.

I have developed and successfully deployed a smart EV charging algorithm that takes care of the whole process using clever mathematics, optimization techniques, and algorithms with heuristics.

Charging electric vehicles (EVs) may be more involved than one thinks, as each electric vehicle has a different charging priority, has a different minimum and maximum charging current, and uses different phases. These phases may be shuffled while the EV charging station's grid connection has limitations, to name a few of the many caveats.

I optimized battery charge rates and times to minimize the battery heating and degradation. This required knowledge of a physical model of a battery, especially its resistance.

The resistance depends on battery chemistry, age, and state of charge, not just the current charge/discharge rate. An optimization model has to be flexible to work not with a fixed physical model but with the specific battery it deals with.

A deep convolutional neural network to detect and classify images of animals into three categories: cats, dogs, and pandas. I created the model, which was an adaptation of a general pre-trained VGG16 model to offset the limited amount of training data available. The model classification accuracy on the validation data was over 97%.

The Math Glass puzzle for Android phones uses simple arithmetic operations to provide an intellectual challenge at various difficulty levels. Primary school students can solve the initial levels. At later levels, one discovers that the human brain is too limited to test all the possible combinations to find the solution; this is where mathematical ingenuity is required to find shortcuts to success. In the end, a diligent player will discover basic tricks of number theory.

A simple Android puzzle accompanied by a character named Sophia. She is a baby mathematician who uses the beauty of mathematics at various levels of difficulty to find out the hidden digit in the number you are thinking of.

Can you figure out how Sophia can guess the hidden digit?

I co-developed a weather service to serve the weather forecast data for a specific location. The service supports crawling data in various formats, interpolation, historical data retrieval, and live streaming.

I forecasted the weather and the associated solar power production, wind power production, site consumption, and other weather-dependent metrics.

Forecasting physical phenomena involved dealing with physical models and concepts and understanding the properties of the electricity components, such as albedo effects, temperature-dependent efficiency, solar cell behavior, Hellman's exponent, wind-power production curve, and others.

I created an algorithm to trade the BTC/USD pair on the FTX cryptocurrency exchange. The algorithm took advantage of low FTX fees to make orders and performed several daily trades. The CAGR in the sandbox environment was positive. However, the real deployment had to be discontinued due to the FTX liquidation.

I contributed to the algorithm for energy trading on the EPEX market. Trading was done to exploit market opportunities and efficiently secure the client's energy resources. The trading consisted of long-term (base-peak contract) and short-term (15-minute energy delivery) trading. The emphasis was not only on profit making but also on risk and penalty minimization while ensuring the smooth supply of energy to the client.

Tokenized raw open text dataset. Trained and fine-tuned a large language model (LLM) on the tokenized dataset. The project adapted the nanoGPT GitHub project from Andrej Karpathy. Worked with models of up to one billion parameters.

Wrote an encryption and a decryption tool that uses LLMs to generate ciphertext indistinguishable from normal text. The tools are compatible with any LLM compatible with Hugging Face Transformers, including Llama 3, Qwen 2.5, and Mistral models.