Yehor Brodskiy

Summary

I'm a Senior Developer from Kyiv, Ukraine. I take the IT problems that don't really fit a SaaS model: backends, infrastructure, streaming pipelines, embeddable forms, on-site builds - and handle them through from the initial interaction to the handover. I'm currently a Senior Developer at The Parker Avery Group, working on enterprise systems for US retail and CPG.

Experience

1. Senior Developer, The Parker Avery Group

   Mar 2022 - Present

   Full-time

   Enterprise systems integration for US retail and CPG: the Unified Demand Intelligence platform and Ask Udi.

   • Designed and solely built an OpenAI-compatible Batch API gateway in Rust (Axum, Tokio, SQLite) in front of a self-hosted vLLM server: durable job state that survives restarts, a two-class scheduler that prioritizes synchronous traffic while never starving batch work, schema-validated responses with a bounded re-ask budget, and retry-with-backoff that honors Retry-After.
   • Deployed and load-tested vLLM serving a quantized open-weight model (Gemma, 12B, 4-bit and FP8 variants) on a single 24 GB GPU. Tuned KV-cache dtype, sequence concurrency, GPU memory utilization and the structured-output backend to reach about 335 output tokens per second at 16-way concurrency with zero request preemptions.
   • Main author of the retrieval layer of a retrieval-augmented, natural-language-to-SQL assistant (LangGraph, LangChain, OpenAI-compatible models) for a large enterprise retail analytics platform. Built a hybrid retriever (pgvector cosine similarity plus PostgreSQL full-text search) that pulls the most relevant prior question and SQL examples into the prompt, in-database embedding generation with sentence-transformers, schema-aware resolution of free-text mentions against an embedded catalog of dimension values over a complicated multi-table snowflake schema, and confidence-driven routing to a cheaper or stronger model by retrieval score.
   • Served open-weight LLMs on Kubernetes and across multiple GPUs over NVLink.
   • Added first-class self-hosted inference to a multi-provider LLM pipeline (local vLLM alongside hosted OpenAI, Gemini, and Cerebras-served gpt-oss), behind one OpenAI-compatible interface with shared retry, timeout and structured-output handling.
   • Delivered a fully air-gapped, self-hosted deployment for a confidential defense-sector client (name withheld under NDA), with no public inbound exposure.
   • Designed and built a GraphQL query gateway that stitches several internal services into one endpoint, with a separate real-time channel for live updates.
   • Ran a capability and throughput study, with load, concurrency and wall-clock gates, to make a self-hosted-versus-cloud go / no-go call for a production batch workload.
   • Diagnosed and worked around an upstream vLLM structured-output bug by pinning the xgrammar backend, documented with root cause and trade-offs.
   • Hardening and observability: bearer-token authentication with constant-time comparison, loopback-only exposure, Prometheus metrics and structured JSON logging.
   • Works inside a mature engineering culture: a test suite, a CI/CD pipeline that builds and deploys the same way each time, and tracing on the AI path.

   ProjectsThe Parker Avery GroupUnified Demand Intelligence

2. Senior Technical Consultant, Mobilauto

   2022 - 2024

   Contract

   I'd already moved to The Parker Avery Group when Mobilauto came back to life. I owned a lot of its parts, so I pretty much got pulled into a 2-year knowledge transfer.

3. Full Stack Engineer / DevOps (incl. SRE), Team Lead, Mobilauto

   Jan 2020 - Mar 2022

   Full-time

   Ran the taxi platform end to end: the Cabinet24 CRM, the Express Soft dispatch platform, Mobilauto Web, the gateway and the partner bridge. prpcow and rust-wasmpack-loader came out of this work as open source. It ended in spring 2022, with the full-scale invasion.

   • Built a program set for taxi services on Kubernetes (k0s) for deployment orchestration. Node.js backend (including worker threads), Kafka for driver-point delivery, Redis for cache, history and pub/sub, PostgreSQL, React frontend, full Asterisk and A2Billing telephony integration.
   • Designed low-cost infrastructure on Kubernetes (k0s) for high availability of every critical service.
   • Led a team of six.
   • Built a large-scale k0s environment supporting multiple deployment options, including automated setups in dedicated client VMs.
   • Built a CRM (client base, task board, calls, mailbox, a file cloud on MinIO and S3 on bare-metal, and an in-browser SIP phone) with full Asterisk and A2Billing integration.

   ProjectsMobilautoCabinet24Express Softprpcowrust-wasmpack-loader

4. Full Stack Engineer / DevOps (incl. SRE), Qberry

   Apr 2019 - Dec 2019

   Full-time

   My first salaried job, at an outsourcing firm. Production reliability and product work.

   • Built many customer-requested environments: Docker-based, bare-metal, several cloud platforms, and Kubernetes clusters including air-gapped setups.
   • Customer projects (React and Vue frontends, Node.js including worker threads, MariaDB, MySQL and MongoDB), plus support for bare-metal servers including PowerDNS.
   • Built gamification add-ons (Node and RabbitMQ on the backend, React on the frontend).
   • Led 2 or more developers.

   ProjectsQberryPlaton P2P

5. Full Stack Engineer / DevOps (incl. SRE), Tech Lead, Cropman.io

   Jan 2019 - Apr 2019

   Own product

   A multi-image editor right in the browser. Built it with classmates from DAN.IT and sold it as a product.

   • An online batch photo editor (layers, crop, filters, watermarks). Node and Sharp (libvips) with Firebase for auth and storage, MongoDB, and a React frontend.
   • Used Firebase Cloud Functions as compute for user requests, with everything else on Docker Compose.
   • Led a team of four.

   ProjectsCropman.io

6. Freelance, Self-taught builds

   2015 - 2018

   Freelance

   Taught myself the stack and put my first sites into production as freelance work.

Skills

• LanguagesTypeScript / JavaScript, Python, Rust, Go, Java, C#
• RuntimesBun, Node, WASM in the browser, Edge runtimes
• FrontendNext.js, React, Tailwind CSS, Angular, Vue, Nuxt.js
• Backend & messagingNode, prpcow, REST, RPC, GraphQL, WebSockets, gRPC, Protobuf, Spring, Microservices
• OrchestrationKubernetes, K8s on bare metal (Talos, k0s), ArgoCD, Devtron, Custom CSI / Operators, k3s, Docker / Podman, CI/CD automation, Terraform, Ansible
• MultimediaHLS, Streaming pipelines, nginx-vod-module
• Auth & identityKeycloak, Zitadel, OIDC / OAuth2, SSO / IdP federation
• CMSStrapi, Custom headless
• Data & AIClickHouse, pgvector, qdrant, OpenAI APIs, Langfuse, LangFlow, LangGraph, RAG / embeddings, LangChain, Apache Spark, DuckDB
• Queues & gatewaysKafka, Valkey / Redis, Higress, RabbitMQ, KubeMQ
• ObservabilitySigNoz, OpenReplay, Microsoft Clarity, Datadog, Prometheus, Grafana
• Cloud & hostingSelf-hosted infrastructure, Bare metal, Major clouds, Private, air-gapped and hybrid environments, EDGE computing
• Realtime & telephonyAsterisk, A2Billing, WebRTC softphones, Custom TURN / STUN with strong auth
• DatabasesPostgreSQL, MySQL / MariaDB, SQL Server, Oracle DB, MongoDB, Cassandra, Couchbase, SQLite
• StorageS3 / MinIO, GCS, Azure Blob, B2, NFS / SMB, Ceph
• LLM servingvLLM, Ollama, Quantization (w4a16, FP8), Multi-GPU serving over NVLink, OpenAI-compatible APIs (Chat and Batch), Throughput and latency load-testing
• SecurityNetwork hardening, VPN-only and no-public-inbound topologies, Secrets management, Vulnerability scanning, Compliance controls

Education

• Kyiv Aviation Institute (National Aviation University)

  2021 - 2025

  Bachelor's, IoT

  Faculty of Electronics, Robotics, Monitoring and IoT Technologies. Average mark 89/100. Qualification work: Network Protocol for Remote Function Calls (RPC-like) as the primary seamless connection between EDGE devices in IoT.

• Kyiv Aviation Institute (National Aviation University)

  2025 - 2026

  Master's, Electronics, Robotics and IoT

  Faculty of Air navigation, Electronics and Telecommunications.

• DAN.IT Education

  2019 - 2021

  Full stack course

  Average mark 80/100.

Languages

• UkrainianNative
• EnglishC1/C2
• RussianNative