Dénes Tornyi

An SQL incremental parser and enforcer which guarantees grammatical and semantical correctness for a subset of the language without any post-validation. The performance-heavy operations are implemented in a Rust kernel and exposed to Python through PyO3, in which the light operations are written (expanding the grammatical tree of SQL and managing the semantic context).

The parser is composed of the following layers: constants as token-level DFAs built by a single Aho-Corasick tree of the vocabulary; regular expressions as lazy DFAs over Brzozowski derivatives; exceptions of reserved keywords as differences of parallel DFAs; the grammar as an infinitely growing graph determinized on the fly as each DFA at acceptance expands into routes; and a semantic state machine which depends on a GF(2) boolean algebra solver and a graph-based table import and join solver for consistency enforcement (of fields, tables and their relations). The state machine supports query stacks, grouping, ordering, filtering and functions.

The project initially used SageMath, but later the subset of its Graph and BooleanPolynomialRing modules were ported to Rust to ensure WASM compatibility.

A Horizon Europe research project dedicated to investigating executive aggrandizement in European democracies, examining whether members of parliament act for their own benefit or represent the regions they speak for. Designed and implemented an automated data pipeline backed by a centralized SQL database to collect and organize parliamentary records across multiple election cycles: politician details and mandates, historical faction transitions, legislative bills, roll-call voting records, plenary speeches and governmental decrees.

Returned to Nokia with the goal of continuing my research on LLMs and RAG systems in Riku Kaura-Aho's team. Established Project Octopus as a centralized project space for publishing and inner-sourcing NLP solutions across the company. Collaborated with various departments to streamline and reduce the internal chaos within R&D operations.

Conducted research on a pipeline architecture inside PostgreSQL using PGRX to write extensions in Rust that minimized HTTP/1.1/SQL communication overhead between services. The design integrated chunking, embedding and retrieval into a single monolithic database process using fastembed and pgvector.

Designed and implemented a microservice-based data warehouse megaproject along with a retrieval-augmented generation (RAG) pipeline composed of a dozen independent Go microservices that were coordinated through a shared PostgreSQL database. The most notable challenges were converting every incoming document to Markdown; writing a reusable TF-IDF library which validated documents and produced their sparse embeddings; implementing a chunking service which broke up text along natural language separators (between paragraphs, sentences or words); and figuring out the right architecture to universally support any dense embedding model, any vector database and any third-party front-end UI. The entire solution was deployed via Docker and used HTTP/1.1 and SQL for inter-service communication.

Designed and implemented a repository of reusable Go packages for cross-project code sharing. The most notable packages were archive, which solved the algorithmically complex challenge of orchestrating multiple concurrent streams into a single stream, enabling parallel log collection from K8s nodes and on-the-fly merging of the incoming log streams into a single tape archive; and httpdump, which provided a transparent HTTP transport middleware for capturing, dumping and replaying HTTP network communication.

Conducted research for Alonso Silva from Bell Labs on predictor machines for enforcing grammatical and semantic constraints on domain-specific languages.

Designed and implemented a matchmaking system to connect local university graduates with local companies (in the Netherlands) by processing company websites and student theses to construct an NxN similarity matrix that described their likelihood of sharing a mutual interest.

Beyond the main project, conducted code reviews, refined the company's CI/CD pipeline and refactored the existing user management architecture to support not only roles but also fine-grained per-user permissions by forking and pruning the AuthPermissions.AspNetCore library.

Designed and implemented (in Go using the K8s Operator SDK) an update orchestrator and an update receiver service to handle firmware rollouts for customer-owned hardware. Developed a MinIO-based firmware storage service with synchronization between regional data centers along with the implementation of HTTP range headers on S3 objects and client-side tools for managing firmware entries.

Contributed to the development of OTT client-side applications for webOS, Tizen and web platforms (for Yle, OSN, NRK, ProSieben, Deutsche Telekom and HBO) using JavaScript with an internally developed library and ReactJS. At peak, supported 10 different applications with bug fixes as a support engineer.

Designed and implemented a real-time third-party API validator middleware between the application and the network layer; a containerized CI/CD toolset for Tizen and webOS; and an internal service operated inside Amazon AWS to synchronize Salesforce time tracking with the BambooHR calendar.

Contributed to the development of an online second-hand marketplace using TypeScript and AngularJS on the client side, C# and ASP.NET Web API 4 on the server side and Entity Framework 6 for database management with Apache Lucene indexing.

Designed and implemented a Facebook Messenger-like service using WebRTC and SignalR over the existing architecture: enabling audio/video calls, text messages along with their message history and the synchronization of ongoing conversation tabs across devices.