Tamas – LLM, AWS, OpenAI

An AI assistant that lets engineers ask natural-language questions about a large, fast-changing codebase instead of digging through files and history by hand. It splits the code into chunks, embeds them in a vector store, and answers questions through an LLM over the retrieved code. It also supports queries against older versions of the repository.

• Built and shipped the assistant end-to-end on a Python/FastAPI backend with a small OpenAI model; it ran in production for daily use by the engineering team.
• Designed version-aware indexing pinned to commit SHAs, with tombstoning for deleted code, so queries about years-old versions still return accurate results.
• Implemented blue-green index switching with atomic cutover, so a failed reindex never exposes partial data to users.
• Replaced nightly rebuilds with incremental reindexing triggered from CI, re-processing only the files changed in each commit.
• Chose WebSockets over REST for the indexing path to handle unpredictable indexing durations.
• Defined retrieval evaluation (recall@k, precision, MRR) and per-request logging of chunk IDs, scores, index and model versions, and latency.

A document intelligence system for a German government organisation that had to process large volumes of mixed-format material by hand. It ingests text, scanned documents, images and video, extracts their content through OCR and multimodal pipelines, and lets staff search and query it in natural language. A multi-agent LLM layer routes each request to specialised agents for retrieval, extraction and answer generation. A fine-tuned small language model runs on a private server, so sensitive data never leaves the client’s infrastructure.

• Led the system end-to-end as technical lead, from technology selection and resource planning through production rollout, observability and evaluation.
• Designed the RAG and multi-agent architecture with LangChain and Pinecone, including ingestion pipelines for text, OCR’d scans, images and video.
• Fine-tuned and deployed a small language model on private infrastructure to meet the client’s data-residency requirements.
• Wrote the architecture and implementation documentation used for the client’s approval process.
• Set coding guidelines, ran PR reviews of AI-generated code with a focus on design-level flaws, and mentored junior and senior engineers on the team.

A cryptocurrency platform serving both end users and internal operations teams. Resolving operational issues used to require manual API calls and direct database fixes. The platform includes customer-facing web applications, backend APIs, an internal approval portal, and admin tooling for data import/export, notifications and reporting.

• Built a hybrid Python and Node.js backend with a maker-checker approval workflow, eliminating manual API calls and direct database interventions for operational issue resolution.
• Developed production Next.js/TypeScript applications and reusable admin modules for data import/export, notifications and reporting.
• Added traceable workflow state and kept roughly 80% Jest/Cypress branch coverage across critical flows, speeding up releases and debugging.
• Coordinated API contracts and release dependencies across backend, frontend, QA and DevOps teams.

An internal operations platform for a national railway. Teams there managed daily tasks through manual, disconnected processes. It provides reporting dashboards, automated workflow notifications, and integrations that move operational data between enterprise systems.

• Delivered API-connected data import/export, automated notifications and reporting dashboards, cutting manual operational work by more than 25%.
• Built backend services in Java and C#/.NET plus integration pipelines connecting enterprise applications and operational data flows.
• Developed React and Angular frontends for the dashboards and daily workflow tools.

A suite of web applications for small and mid-sized businesses running online stores and day-to-day operations. It brings payments, storefront management, business tools and real-time sales analytics into one place.

• Built real-time ETL/streaming pipelines that cut analytics latency by 50% across customer-facing products.
• Integrated Stripe, PayPal, Shopify, Google Workspace and Office 365 APIs for payments, commerce and business workflows.
• Shipped production Laravel/MySQL and Vue.js applications in a fully remote startup team.
• Defined technical scope, timelines and integration requirements directly with project managers.