Sovereign software engineer building AI, blockchain, and cybersecurity infrastructure from Türkiye. Founder and lead engineer at eCloud Yazılım Teknolojileri.
I work where these three disciplines intersect — sovereign AI that is KVKK-resident and verifiable, blockchain rails that satisfy real compliance regimes, and security postures that hold under regulatory audit. The work below is the public, measurable proof of that practice.
A 128-billion-parameter sovereign, multimodal large language model designed for Turkish. Released April 2026. Built in Türkiye, evaluated against the world.
| World #1 Turkish reading comprehension | Belebele-TR 87.33 |
| Frontier-level grade-school mathematics | GSM8K 94.62 |
| Frontier-level scientific reasoning | ARC-Challenge 94.88 |
| Upper-mid frontier code generation | HumanEval 84.15 / MBPP 84.82 |
| Lower-mid frontier multitask academic | MMLU 80.10 |
| Honest gap (V4.1 priority) | GPQA-D 37.88 / MMLU-Pro 50.20 |
22 benchmarks · 13,344 real production API calls · Wilson 95% confidence intervals · seed=42 · per-item scored.jsonl open under MIT.
Whitepaper (32 pp, EN + TR) · Model card · Evaluation dataset · Live demo · Benchmark runner · Production API
Sovereign artificial intelligence. Architecture, training, and evaluation of frontier-adjacent large language models with full data residency. Hierarchical Memory (HBM) with Time-Guided Decay, Localised Mixture-of-Experts, Adaptive LoRA+, Selective Layer Collapse, 4-bit symmetric block quantization. Reproducible evaluation as a first-class engineering artefact, not a marketing afterthought.
Blockchain and verifiable infrastructure. GPG-signed training pipelines, content-addressable model checkpoints, and tamper-evident audit trails. The same trust primitives that make a chain auditable also make an AI release auditable — I treat them as one stack.
Cybersecurity and compliance. KVKK-resident hosting, EU AI Act posture, and security review for production AI workloads in public-sector, financial, healthcare, and defence contexts. Threat modelling for prompt injection, data exfiltration, and supply-chain risk in model weights and tokenizer artefacts.
Programme founded 2021. Türkiye's longest-running fully sovereign LLM line.
| Year | Release | Milestone |
|---|---|---|
| 2021 | Programme founded | Sovereign-AI research charter; Turkish-first tokenizer studies; data residency roadmap |
| 2022 | Foundational research | Sovereign training infrastructure; Turkish morphology calibration; first internal LLM prototypes |
| 2023 | AIGENCY V1 | First fully sovereign Turkish LLM milestone (early-access cohort) |
| 2024 | AIGENCY V2 | First public sovereign release; Turkish-NLP baseline established |
| 2025 | AIGENCY V3 | Architecture maturity, 200K+ context, hierarchical memory primitives |
| 2026 | AIGENCY V4 | Multimodal, 128B parameters, frontier-adjacent, transparent evaluation |
| 2026 Q4 | AIGENCY V4.1 | Vision encoder 8B → 16B, GPQA-D 0.55 target, latency p50 ≤ 4 s |
| Artefact | Description | Licence |
|---|---|---|
aigency-v4-whitepaper |
32-page technical whitepaper (EN + TR) | CC BY-ND 4.0 |
aigency-benchmarks |
Reproducibility capsule — 22 benchmarks, 13,344 raw responses, scoring code | MIT |
aigencydev/AIGENCY-V4 |
Hugging Face model card and methodology | Open methodology |
aigencydev/aigency-v4-evaluation |
Evaluation dataset (raw + scored) | MIT |
aigencydev/AIGENCY-V4-Demo |
Live Gradio demo (production API proxy) | Open |
Languages. Python · Rust · Go · C++ · TypeScript · Solidity · Move · Vyper · Bash · SQL
| Layer | Stack |
|---|---|
| Training | PyTorch · JAX · DeepSpeed · FSDP · Megatron-LM · accelerate |
| Inference kernels | Triton · CUDA · Flash-Attention 2/3 · xFormers · bitsandbytes · AWQ · GPTQ |
| Serving | vLLM · TensorRT-LLM · TGI · llama.cpp · Ollama · BentoML |
| Eval and observability | lm-evaluation-harness · lighteval · Weights & Biases · MLflow · Phoenix · Langfuse |
| RAG and vectors | Qdrant · Weaviate · Milvus · pgvector · FAISS · LangChain · LlamaIndex |
| Multimodal | CLIP / SigLIP · Whisper · SAM 2 · Florence-2 · vision encoders · cross-modal projection |
| Agents and tools | MCP (Model Context Protocol) · AIGENCY SDKs · tool-use schemas · structured output |
| Layer | Stack |
|---|---|
| Smart-contract development | Foundry · Hardhat · viem · ethers.js · web3.py |
| EVM and L2s | Ethereum · Optimism · Arbitrum · Base · zkSync · Scroll · Linea · Polygon zkEVM |
| Non-EVM | Solana (Anchor, Rust) · Cosmos SDK · Sui · Aptos |
| Zero-knowledge | Circom · Noir · Halo2 · Plonky3 · RISC Zero · SP1 · snarkjs |
| Standards | ERC-20 · ERC-721 · ERC-1155 · ERC-4337 (account abstraction) · EIP-712 · ERC-6551 |
| Audit and analytics | Slither · Mythril · Echidna · Foundry invariants · The Graph · Dune · Tenderly |
| Verifiable AI | content-addressed checkpoints · on-chain attestation registries · restaking primitives (EigenLayer) |
| Layer | Stack |
|---|---|
| Static and dynamic analysis | Semgrep · CodeQL · Bandit · Burp Suite · OWASP ZAP · nuclei |
| Network and offensive | nmap · Wireshark · Metasploit · Hydra · BloodHound |
| Supply-chain integrity | Sigstore (cosign, rekor) · SLSA · in-toto · Syft · Grype · Trivy · GPG · age |
| Cloud and runtime | Falco · OPA / Gatekeeper · Kyverno · AWS Security Hub · GCP SCC · Azure Defender |
| SIEM and detection | Elastic SIEM · Wazuh · Splunk · Suricata · Zeek |
| Identity and secrets | HashiCorp Vault · OAuth 2.1 · OIDC · SAML · mTLS · WebAuthn |
| AI-specific security | prompt-injection threat models · data-exfiltration controls · tokenizer / weight supply-chain · jailbreak red-teaming |
| Compliance frameworks | KVKK · EU AI Act · GDPR · ISO 27001 · SOC 2 Type II · NIST CSF 2.0 · PCI DSS |
Kubernetes · Terraform · Pulumi · Ansible · Helm · Argo CD · FastAPI · gRPC · Next.js · Postgres · ClickHouse · Redis · Kafka · NATS · Grafana · Prometheus · OpenTelemetry
If you reference AIGENCY V4 in academic or technical work:
@techreport{aigency-v4-2026,
title = {AIGENCY V4: Sovereign, Fully Independent and Multimodal 128B-Parameter AI Architecture},
author = {{eCloud Yazılım Teknolojileri}},
year = {2026},
month = apr,
institution = {eCloud Yazılım Teknolojileri},
url = {https://github.com/ecloud-bh/aigency-v4-whitepaper}
}(Use a UTF-8 BibTeX backend — biber + biblatex — for correct
rendering of Turkish characters. For legacy bibtex, replace
Yazılım with Yaz\i l\i m.)
- Direct. mpolat@e-cloud.web.tr
- AIGENCY team. ai@aigency.dev
- LinkedIn. linkedin.com/in/muhammetpolat
- Web. aigency.dev · e-cloud.web.tr
Open to research collaboration on Turkish NLP, sovereign AI, and KVKK-compliant inference. Enterprise access and partnership inquiries welcome via either address above.
Built in Türkiye, evaluated against the world. © 2026 eCloud Yazılım Teknolojileri.