Mert Sırakaya - Agentic Systems Engineer

Mert Sırakaya

Agentic Systems | Python | OSS

Istanbul, Türkiye | contact@tomris.dev

§ github.com/fswair | ï linkedin.com/in/mert-sirakaya |  tomris.dev | 3 pypi.org/user/swair

Intro

Applied AI Engineer and Python Developer focused on open-source software and reliable LLM-powered

systems. I build typed Python libraries and services, with particular attention to maintainable APIs, observability,

and evaluation. I am especially interested in interoperability between agent runtimes and protocol-based tool-

ing such as ACP. I work with a disciplined testing approach, using pytest and meaningful test coverage to keep

changes safe as projects evolve.

Experience

Freelance Python Engineer Jun 2022 – Present

Independent Client Work Remote

–

Delivered Python APIs, automation, and data workﬂows for clients, including a high-concurrency

FastAPI/Web3 service used by 250+ clients and serving 2M+ edge requests with approximately 400

ms latency and 99.2% uptime.

Open Source Projects

ACP Kit | Maintainer Apr 2026 – Present

–

Maintain an open-source toolkit and monorepo for connecting existing agent implementations to the Agent

Client Protocol.

– Pydantic ACP

exposes existing Pydantic AI agents as ACP servers while preserving their models, tools, de-

pendencies, streaming behavior, and conﬁguration. The monorepo also contains langchain-acp, codex-auth-

helper, and acpremote.

Kedi | Typed language and runtime for LLM orchestration Jul 2025 – Present

–

Privately developing a language and runtime that combines native programming constructs with typed ab-

stractions for context-aware prompts and LLM orchestration.

–

Designed language-level evals, tests, reusable programming primitives, and prompt spans that can be opti-

mized with GEPA using evaluation feedback. Example | Demo

Deck Validator | End-to-end PowerPoint review system Jul 2026

–

Built a review pipeline combining deterministic PPTX/OOXML validation with a Pydantic AI-backed

LLM/VLM judge, Logﬁre tracing, calibration against annotated examples, and a human review workﬂow.

–

Published the redacted implementation, reproducible validation traces, and generated report.

Repository |

Demo | Validation Report

VOWEL | Evaluation tooling Sep 2025 – 2026

–

Previously maintained an evaluation DSL and eval-generation framework built on Pydantic Evals, combining

standard assertions with LLM judge evaluators.

–

Built a Pydantic Monty-backed CodeMode pipeline with feedback-guided exploration and created

codemode-benchmark to compare model pairs across success rate, coverage, runtime, and cost.

Technical Skills

Python: Python 3, object-oriented programming, concurrent programming, type annotations, FastAPI, Pydantic,

package design, and PyPI publishing

Applied AI: Pydantic AI, structured generation, tool calling, LLM integrations, LLM/VLM evaluation, prompt

optimization, DSPy, and GEPA

Open Source and Protocols: ACP, MCP, API design, package maintenance, versioning, protocol adapters, and

technical documentation

Testing and Code Quality: pytest, unit and integration tests, ﬁxtures, mocks, parametrization, test coverage,

Codecov, and GitHub Actions

Evaluation and Observability: Pydantic Evals, LLM judges, benchmark design, calibration, Logﬁre, tracing,

monitoring, and production debugging

Infrastructure and Data: Docker, AWS, Nginx, SQL, SQLite, SQLAlchemy, and MongoDB