Software Architect · Cloud & DevOps · AI

I’m Ali. I architect cloud platforms — and build AI you can verify.

17+ years building payment, banking and trade-finance systems. Today I architect Java platforms on AWS and GCP, run them with GitOps, and research generative AI for regulated environments — agentic, model-agnostic and deterministically verified.

Get in touch → LinkedIn Frankfurt am Main
Ali AminianSoftware Architect · Platform Engineer · AI Engineer

Currently I’m Platform Engineer & Software Architect at Yatta in Frankfurt, working on event-driven integration and platform engineering with Kafka, Temporal and Kubernetes.

Before that I owned the architecture of a payments platform at ZahlungsWERK on GCP, built backend services for Deutsche Telekom’s MagentaTV, worked on trade finance for Deutsche Bank, contributed to AI-driven trade-finance document automation at Complidata, and led a 10+ person trade-finance team. Earlier I delivered banking systems across more than nine banks.

Separately, I run independent generative-AI R&D projects: an EDIFACT integration prototype where LLMs draft partner mappings and the system machine-verifies them, and an on-prem document examination prototype for trade-finance rules. Both use an AI-assisted engineering workflow — Claude Code with custom agents, skills and MCP integrations, driving spec-first development and automated review.

Expertise

Backend. Cloud. DevOps. AI.

Four things I do at senior level.

01
~40%lower latency

Java Backend Engineering

High-throughput services that move money — built on modern Java and Spring.

  • Payments platform on Java 21 & Spring Boot 3 with ~40% lower latency
  • Exactly-once event publishing: Transactional Outbox, Kafka, Avro & Schema Registry
  • Spring Boot → Quarkus migration with ~50% faster startup
Java 17–25Spring Boot 4WebFluxQuarkusKafkaTemporalgRPCPostgreSQL
02
−35%processing time

Cloud Architecture · AWS & GCP

How systems run in the cloud — networking, security, identity and data flow.

  • Service mesh and ingress design with Istio; OAuth2/OIDC, RBAC and fine-grained authorization
  • Payments platform on GKE with Pub/Sub and Debezium (−35% processing time)
  • Change data capture and event streaming with Kafka Connect and Debezium
AWS · EKSGCP · GKE · Pub/SubAzure · AKSIstioOpenFGAGraphQL Federation
03
~50%less manual deployment

DevOps & Platform Engineering

Infrastructure as code, GitOps and observability — releases that are boring on purpose.

  • Terraform and Helm automation that cut manual deployment effort ~50%
  • GitOps CI/CD with ArgoCD and Helm across environments
  • Zero-downtime blue/green and canary releases with Istio
KubernetesTerraformHelmArgoCDOpenTelemetryPrometheusTempoLoki
04
9.5×faster, zero egress

Applied AI

Independent R&D on agentic, model-agnostic LLM systems for regulated domains.

  • On-prem trade-finance document examination with deterministic verification
  • EDIFACT integration prototype with machine-verified LLM mappings
  • Quantized serving, speculative decoding and prefix caching on a single A100
LangGraphRAGMCPClaude APIvLLMOllamaQdrantSpring AIGeminiDocument AI
Domain

Payments · Dunning · SWIFT MT/MX · Trade Finance · Core Banking17+ years across 9+ banks and fintechs

Independent R&D

AI that has to prove its work.

Personal research projects built end to end — local model serving, agent design, retrieval, verification and tests.

Logistics · EDIMachine-verified

AI-native EDIFACT integration prototype

Onboarding a trading partner traditionally means weeks of hand-written mappings. Here an LLM reads a partner’s spec and sample files and drafts the mapping — then the system refuses to trust it. DFDL conformance, field-coverage analysis and dry runs against real-world EDI files decide whether it passes, and a failed check goes back to the model with the exact reason.

  1. Partner spec & samples
  2. LLM drafts mapping
  3. DFDL + coverage checks
  4. Dry run on sample files
  5. Verified mapping
Byte match~99.5%Tests1,900+Java~116k LOCStackJava 25 · Spring Boot 4
Trade financeOn-prem · Zero egress

Trade-finance document examination

Checks letter-of-credit document sets against UCP 600, eUCP and ISBP the way an expert examiner would: invoices, bills of lading and certificates read against the credit and its SWIFT MT700 terms. Multi-agent LangGraph pipeline, local OCR and a local LLM, with deterministic verification on top — every discrepancy structured, rule-referenced and anchored to the page it came from. Rules are editable text, not code, so a change in practice needs no release. Everything runs locally, with zero data egress.

  1. Scan & SWIFT intake
  2. Local OCR
  3. Structured extraction
  4. Agentic rule checks
  5. Deterministic verification
  6. Source-anchored findings
Extraction9.5× fasterThroughput/day500+ casesGPU1× A100Tests~960
Personal projectGo · Kubernetes

Kubernetes traffic-learning tool

A learned twin of a Kubernetes cluster: a read-only recorder learns how microservices behave from real traffic, redacting it inside the cluster, and one binary on the laptop stands in for those services. A developer can run one service in the IDE without touching shared staging.

  1. Read-only recording
  2. In-cluster redaction
  3. Learned behaviour
  4. One local binary
Model-agnosticSwap models without rewrites — vLLM, Ollama or cloud APIs behind one interface.
Agentic by designMulti-agent LangGraph workflows — tools exposed via custom MCP servers, state, and human checkpoints.
Verified, not trustedDeterministic checks on every AI output before it counts.
Private by defaultOn-prem serving, zero data egress, full audit trail.
MeasuredEvals and automated tests — not vibes.

Local & self-hosted LLMs

  • Serving open-weight models with vLLM and Ollama, fully offline
  • Quantized weights, speculative decoding and prefix caching for throughput
  • GPU sizing and batching — throughput of 500+ multi-document cases a day on one A100
  • One interface over local and cloud models, so either can be swapped in
vLLMOllamaQuantizationSpeculative decodingPrefix cachingGPU tuning

Retrieval & context engineering

  • RAG over scanned, messy documents — layout-aware chunking and OCR clean-up
  • Hybrid search with reranking on Qdrant, cited back to the source page
  • Structured outputs against a JSON schema instead of free text
  • Prompt and few-shot design that survives model changes
RAGQdrantHybrid searchRerankingStructured outputPrompt engineering

Agents, tools & evaluation

  • Multi-agent LangGraph workflows with state, retries and human checkpoints
  • Custom MCP servers exposing internal tools and data to agents
  • Golden-set evals and regression tests in CI — ~960 and 1,900+ automated tests
  • AI-assisted delivery with Claude Code: custom agents, skills, spec-first development
LangGraphLangChainMCPClaude APIClaude CodeSpring AILLM evaluation
Work

Where I’ve been

Full history on LinkedIn →

Yatta GmbH

Platform Engineer & Software Architect

2025 — Present

Independent Generative AI R&D Projects

Banking compliance & logistics EDI

2026 — Present

ZahlungsWERK GmbH

Architect & Senior Backend Developer

2023 — 2025

Complidata

Senior Backend Developer (freelance, part-time) — TradeSpeed

2024

axxessio · Deutsche Telekom

Senior Software Engineer

2019 — 2023

GFT · Deutsche Bank

Functional Analyst, Trade Finance

2019

Dotin

Trade Finance Team Lead

2014 — 2019

Banking & FinTech

Consultant & Developer

2006 — 2014
Education

B.Sc. Computer Science (Hardware), Azad UniversityEnglish (fluent) · German (B1) · Persian (native)

Contact

Want to get in touch?

Email is best — I read everything.

aminian.ali@gmail.com