Yauheni Padniuk — AI Architect & AI Software Engineer
AI Architect & AI Software Engineer based in Kraków, Poland, with 12+ years in platform engineering. I design and ship production LLM systems — agents, MCP servers, model routing and cost controls — and choose models based on benchmark results, not hype.
About
Platform Architect / SRE Tech Lead at ScienceSoft: multi-cluster Kubernetes for 20K+ edge devices across 50+ microservices with 99.9% uptime, LLM-serving infrastructure (vLLM, LiteLLM, MLflow) and AI coding-agent adoption across the platform team. Designed and shipped a production AI SaaS end to end — LLM platform, agents, MCP server and eval harnesses — with 2,000+ Claude Code and 190+ Codex sessions orchestrating 15K+ sub-agent runs.
AI Engineering
- Evals, benchmarks & data-driven decisions: golden corpora with real-user holdout, P/R/F1 and quantity accuracy, structural gates vs semantic trends. Picked the import model on eval (89% vs 76% quantity accuracy at equal cost), cut physics-model water-activity RMSE 0.117 → 0.081 on a 404-recipe measured corpus, routed the assistant to a cheap tier after measuring tiers on a corpus.
- AI cost engineering: cost per successful task as the model-selection KPI, micro-USD reserve-and-settle ledger, CI spend cut ~70% from a 5,343-run attribution, agent-facing check output cut from ~180K tokens to ~25 lines, MCP tool-result context cut ~75%.
- Production LLM platform: single LiteLLM egress to Anthropic and OpenAI, server-side model routing, durable runs with exactly-once-safe settlement.
- Agents & tool calling: bounded agent loops with deterministic tool steps, typed outcomes, plan-once import agent with byte-identical execution, human-in-the-loop catalog agent.
- MCP server: remote MCP on the official Go SDK — Streamable HTTP, OAuth 2.1 with PKCE, per-tool scopes, rate limits, 30 tools and interactive widgets for Claude and ChatGPT.
- LLM infrastructure: on-prem LLM appliance control plane (vLLM, LiteLLM) and MLOps platform (MLflow, agent hosting) on dedicated EKS, captured in 15 ADRs.
- AI-native delivery & governance: parallel worktrees, sub-agent fan-out, agent teams; hooks, rules and skills; production read-only guardrails for coding agents; Claude PR review across 6 repos; mentoring engineers on Claude Code and Codex.
Tech Stack
AI Engineering & LLM Systems: Anthropic & OpenAI APIs, LiteLLM, tool calling and agent loops, structured outputs, MCP (OAuth 2.1), RAG (pgvector + lexical), prompt caching, LLM cost governance.
Evals, Benchmarks & KPIs: golden corpora with holdout, P/R/F1, RMSE and bias analysis, regression gates, cost per successful task, DORA, SLO / error budgets, data-backed roadmaps.
AI Infrastructure & AI-native Dev: vLLM, MLflow, LLM gateways, Claude Code, Codex, multi-agent orchestration, hooks / skills / rules, agent guardrails.
Platform & Cloud: AWS (EKS, Lambda, S3, RDS), Azure, GCP, Multi-cluster Kubernetes, Hybrid & Air-gapped solutions, Service Mesh, SOC2 & GDPR compliance.
DevOps: Kubernetes, Docker, Helm, Terraform, Ansible, GitOps, ArgoCD, CI/CD, Istio, Linkerd.
SRE & Observability: SLI/SLO/SLA, Error Budgets, Prometheus, Grafana, Loki, Jaeger, Chaos Engineering.
Languages & Data: Go, Python, TypeScript, Bash, PostgreSQL, pgvector, Redis, Kafka, RabbitMQ, Elasticsearch.
Experience
Platform Architect / SRE Tech Lead — ScienceSoft (2021–Present)
Leading architecture of a multi-cluster Kubernetes platform serving 20K+ edge devices — 10,000+ RPS, 50+ microservices, 99.9% uptime. Owned DORA and SLO/error-budget KPIs: weekly → 50+ deploys/day, MTTR −80%. Showed from git history that ~50 developer changes in 3 months reached production through only 3 manual SRE promotions (a 16:1 queue); leadership approved the fix and I built the multi-tenant Developer Platform, where a new app onboards in about an hour with policy guardrails by default. Architected LLM-serving infrastructure (vLLM, LiteLLM, MLflow) across 15 ADRs and led AI coding-agent adoption with guardrails and mentoring. Designed air-gapped K8s for regulated industries ($2M+ pipeline). Cut infra costs by 20% ($250K+/year).
DevOps Lead / Platform Engineer — ScienceSoft (2019–2021)
Led team of 12 engineers building real-time media streaming platform — 10,000+ RPS with GPU-accelerated live video transcoding. Migrated monolith to containerized microservices, cutting deployment time by 80%.
DevOps / Backend Engineer — ScienceSoft (2013–2019)
Built ETL pipelines processing 1M+ records/day, monitoring systems for 200+ services, and HIPAA-compliant infrastructure.
Key Achievements
- Data-backed platform strategy: ~50 developer changes queued behind 3 manual SRE promotions (16:1) — initiative approved, new apps now onboard in ~1 hour
- Eval-driven AI decisions: 89% vs 76% import quantity accuracy, water-activity RMSE 0.117 → 0.081, CI spend ~70% lower
- Production AI platform built with agents: 4,100+ commits and 3,100+ merged PRs in its first ten months
- 99.9% uptime for multi-cluster platform with 20K+ edge devices
- $250K+ annual infrastructure cost savings
- 80% reduction in mean time to recovery (MTTR)
- 60% reduction in operational tickets via self-service provisioning tooling
- $2M+ pipeline from air-gapped K8s solutions for regulated industries
Projects
- Production AI SaaS — AI-native B2B SaaS for scientific food formulation, designed and built end to end with coding agents: LLM platform, agents, MCP server, eval harnesses.
- AI-assisted Terraform Plan Review — opt-in LLM review of Terraform plans in a Go Terragrunt CLI, with masked allowlisted export, secret withholding and advisory-only results.
- Agent Trace Cost Estimator — deterministic parser and cheap-model labeling of coding-agent traces into a token and cost estimation catalog (963 sessions, ~57B tokens).
- Disaster Recovery & Data Resilience — DR strategy for 5 critical data stores with RPO <30min and RTO <2h.
- Multi-Cluster Migration — Migration of 20+ customer clusters with zero data loss.
- Infrastructure Load Testing Framework — Automated load testing supporting 10,000+ RPS benchmarking and chaos engineering.
Contact
View traditional CV