Building the GPU orchestration, training, and evaluation platform for fine-tuning and serving open-weight large language models at scale.

• Build and maintain a distributed ML orchestrator in Rust that coordinates fleets of GPU workers for training, evaluation, and inference of open-weight LLMs.
• Engineer the Python/PyTorch training and evaluation platform powering supervised fine-tuning (SFT), preference and metric RLHF, and reinforcement learning recipes for open-weight models.
• Built dataset management for ingestion, versioning, and curation of training and evaluation datasets.
• Designed and shipped the Python training SDK and public API used by customers to run and manage their training and evaluation jobs.
• Contributed full-stack, building product features end to end with a React and GraphQL frontend on top of the platform APIs.
• Operate distributed systems on Kubernetes at scale, including gang-scheduled GPU jobs and high-performance networking.
• Drive performance engineering across the stack, from a custom high-performance RPC protocol to GPU memory management and low-level optimizations.
• Developed the sandboxing component for safe execution of agentic and training workloads.
• Tech stack: Rust (tokio, axum, SeaORM, PyO3), eBpf, Python, React, GraphQL, ClickHouse, PostgreSQL, Redis, Kubernetes, AWS, OpenTelemetry

Built an AI-powered SaaS that helps major construction stakeholders analyze, understand, and identify risks in large construction projects.

• Led engineering for the core AI platform as Engineering Manager while remaining hands-on across the full stack.
• Built document-processing pipelines for extraction, OCR, classification, Retrieval-Augmented Generation (RAG), annotations, semantic search, and summarization.
• Deployed Large Language Models (LLMs: Llama 3, Qwen) and BERT models on Azure; performed LLM fine-tuning (RLHF, RLAIF), synthetic data generation, and model evaluation and selection using Python (PyTorch, LangChain).
• Re-architected the platform to improve scalability and maintainability.
• Tech stack: Rust, Python, VueJS, Postgres, Elasticsearch, Kafka, Kubernetes on Azure, MLFlow, Weights & Biases

Joined an early-stage startup developing a live video chat solution with white-labelling, automatic moderation, analytics, and billing. Product used by luckycrush.live and other social and gaming platforms.

• Scaled the platform to 15M+ users and €1.2M+ monthly revenue, culminating in a successful acquisition by TrafficPartner.
• Owned full-stack development, DevOps, technical support, and team hiring and onboarding as the founding technical lead.
• Built real-time, low-latency network systems in Rust (QUIC/HTTP3, WebRTC, HLS media streaming) powering live video chat at scale.
• Tech stack: Rust, TypeScript, React Native, WASM, cross-compilation, NATS, MongoDB, PostgreSQL, Google Cloud, TimescaleDB

Developed a virtual and physical events management solution from scratch. Led a team of 8 developers.

• Built a video conferencing system using WebRTC and DASH/HLS for virtual events, event registration, and management
• Implemented CI/CD pipelines
• Architected microservices and event-driven systems on Kubernetes
• Mentored junior engineers and collaborated on technical roadmap with CTO
• Built backend services with Nest.js (TypeScript) and set up end-to-end testing with Cypress
• Tech stack: TypeScript, Nest.js, Cypress, Rust, nginx/openresty, Kubernetes, CouchDB, Kafka, Redis

Built 'test as a service' solution (similar to SauceLabs) providing resources for automated end-user testing with VMs, browsers, OS choices, mobile device emulators, and load testing features.

• Migrated monolithic application to distributed, fault-tolerant microservices architecture
• Integrated performance engineering, resources orchestration, virtualization, distributed tracing and monitoring tools
• Tech stack: C#, JavaScript, Docker, cloud-init, ELK, Azure, Postgres

A startup helping clients improve online advertising and pre-targeting through efficient recommendation algorithms.

• Implemented and maintained recommendation algorithms and Python data/analytics processing pipelines
• Focused on performance engineering and optimization
• Tech stack: C#, Python, TypeScript, Docker, MongoDB, ElasticSearch, Spark, Google Cloud

Contributed to building solutions for programmatic advertising.

• Implemented a high-performance Real-Time Bidding (RTB) engine with a reactive, event-driven architecture.
• Built time-series forecasting for inventory optimization and optimized ad targeting and ad campaign performance.
• Tech stack: C#, JavaScript, Aerospike, SQLServer, Hadoop

Part of a 25-developer team building solutions for asset management, accounting, and finance.

• Developed core components for authentication, caching, security and frontend architecture
• Built workflow engine for managing financial instruments lifecycle
• Led migration from monolithic to microservices architecture
• Tech stack: C#, C++, MarkLogic, Redis, SQLServer, Angular 2

Part of an 8-developer team building low-latency automated trading engines for bond markets.

• Implemented connectors to financial markets
• Built order negotiation systems and Request for Quote (RFQ) workflows
• Tech stack: C++, C#, SQL, FIX, ZeroMQ, ION Platform, Low level TCP/UDP

Part of a 5-developer team building a grid-computing solution for financial risk indicators on large portfolios.

• Designed pricing API and data model for financial instruments
• Integrated and wrapped legacy C++/C computation libraries
• Implemented core algorithms for risk computation engine
• Tech stack: C++, OpenBLAS, C#, SQLServer, Protobuf, Apache Cassandra, IBM Platform Symphony

Maintainer of Sayiir (github.com/sayiir/sayiir, 60+ stars), an embeddable durable workflow engine in Rust with Python (PyO3) and Node.js bindings. A lightweight alternative to Temporal and Airflow.

• Durable execution via checkpointing, with crash recovery.
• Tech stack: Rust, PyO3, Node.js, distributed systems