Roman Yastrebov

AI infrastructure and LLM products

• Designed an AI assistant for a Flutter mobile app on a custom agent harness: a tool-calling loop with tools over user data, web search, streamed answers, memory and parsing of uploaded documents. Models are connected through LiteLLM, so the provider changes without code edits and a failed request falls back to a backup model.
• Took the assistant to production: daily quotas, an answer deadline, token accounting and a trust boundary for tool results against prompt injection. Prompts are versioned with the code and updated without an app release.
• Deployed an LLM gateway on LiteLLM: one OpenAI-compatible API for cloud and self-hosted models, per-project keys and budgets, fallbacks between providers.
• Ran open-weight models on my own servers for tasks where data cannot go to external APIs: LLMs on vLLM and Ollama, speech recognition with Whisper and meeting diarization.
• Built RAG search over documents on PostgreSQL with pgvector, with hybrid search and an eval set of questions I use to check answer quality.

Underdogi

A system that recognizes PDF protocols of sports competitions with LLMs and computes ratings of Kyokushin karate fighters.

• Lead the project team: assign tasks, review code, choose models and how recognition quality is measured.
• Built an eval bench and used it to pick models for recognizing 30k pages of protocols. Accuracy went from 53% to 96% at about $0.01 per page.
• Designed a new recognition engine. The model extracts the bracket structure and code rebuilds the tournament logic; disputed pages go to a stronger model, and spending is capped.
• Sped up the rating page from 10 seconds to 60 ms with precomputed snapshots.
• Added validation and deduplication of results, a data snapshot and automatic rollback on failure to the weekly run.

Stack: Python, FastAPI, OpenAI API, Gemini API, PHP 8.4, Symfony 7.4, MySQL, Nuxt 3, Docker.

GigWork

An HRM platform for staff outsourcing. Several PHP/Symfony services and Go microservices.

• Owned the platform architecture. Designed new modules, agreed technical decisions with the team and introduced architecture reviews for major changes.
• Wrote Go microservices for the high-load parts of the platform: RabbitMQ event handlers and services with gRPC APIs.
• Moved mass accruals and payouts to contractors to asynchronous batch processing so that peaks do not block the API.
• Designed piece-rate pay across the whole process: order, assigning contractors, reconciliation, payouts, reporting.
• Proposed and carried out merging the customer portal and HRM into one service. Duplicated code went away and new features started shipping about 25% faster.
• Upgraded the services from PHP 7.4 and Symfony 5.4 to PHP 8.3 and Symfony 6.4 and covered the whole customer portal API with functional tests.

Stack: Go, gRPC, PHP 8.3, Symfony 6.4, PostgreSQL, Redis, RabbitMQ, Graylog, GitLab CI/CD.

Go-Base Platform

An ERP/WMS warehouse system with RFID. Heavy legacy: PHP 7.0, a custom framework on Zend Framework 1, 58 business modules.

• Moved the team to agentic development with Claude Code; tasks started getting done 40% faster.
• Introduced a spec-driven process and automated the routine around tasks: agents create a branch from a YouTrack task, open a merge request and update the status on their own.
• Migrated the codebase from Zend Framework 1 to Symfony components with no regressions for neighbouring teams.

Stack: PHP, Zend Framework 1, Symfony, MariaDB, PostgreSQL, RabbitMQ, gRPC, Claude Code, MCP.

GigWork Exchange

A request exchange for outsourcing companies in the CIS. A product built from scratch.

• Led a team of 8: 4 backend, 2 frontend, 2 QA.
• Laid down the architecture on DDD and Hexagonal, with module boundaries checked by Deptrac.
• Set up TypeScript model generation from OpenAPI: the frontend received contracts automatically, without manual syncing.

Stack: PHP, Symfony, PostgreSQL, Deptrac, Psalm, PHPStan, OpenAPI, GitLab CI/CD.

Too Many Gifts

A B2B store of corporate gifts with over a million products in the catalogue. A team of 4.

• Designed the platform from scratch: the main backend and an import microservice connected through RabbitMQ.
• Built faceted catalogue search on Elasticsearch; filters and counters are computed with aggregations without database queries.
• Designed the import from 15+ suppliers with different protocols (REST, SOAP, FTP, XML feeds) and gigabytes of data. Parsers stream within bounded memory, synchronization is incremental, and a new supplier is added with a single parser.
• Relaunched the live store on the new platform without losing SEO.

Stack: PHP 8.1, Symfony 5.4, PostgreSQL, Elasticsearch, Redis, RabbitMQ, S3, Docker.