Reza Esfahani (Mohammadreza Taleghani Esfahani) is a Senior Software Engineer in Stuttgart, Germany, working on resilient distributed systems, cloud platforms and backend architecture.
Senior Software Engineer
REZA Esfahani, Senior Software Engineer
Backend, observability and FinOps for distributed cloud systems. Currently at Exxeta in Stuttgart.
A few engineering problems I've worked on, what I chose, and what it cost.
Exxeta
AI for FinOps with MCP
Problem
Getting an answer about cloud cost, usage or savings meant asking an engineer or a cloud engineer. Management and finance had to go to the tools, or to the people who run them.
Decision
Bring the tools to management, not management to the tools: a conversational interface in Teams, next to a Power BI FinOps dashboard, instead of one more tool to learn. Built as a proof of concept first.
Outcome
Microsoft Teams Copilot can query cloud account costs, usage metrics and optimization opportunities, so management and finance get answers without asking an engineer. Built as a consultant and presented to the customer, who is now considering it. It sits alongside a FinOps Hub concept I designed on Azure FinOps tools.
MCPClaudeAzureCopilot StudioPower BIExxeta
Central telemetry platform
Problem
One team had no monitoring and read logs on VMs. Another ran Grafana on a VM that could not scale horizontally. Service names differed between logs, metrics and traces, so the three could not be connected.
Decision
Service names are set at the source, not rewritten in a collector, so consistency is built in and each service follows one convention, with no prefixes.
Outcome
One Grafana stack on Amazon EKS for the teams, where logs, metrics and traces share the same service name and can be followed from one signal to the next.
GrafanaLokiTempoAmazon EKSTIER Mobility
Cutting cloud and observability cost
Problem
Cloud usage and observability tooling carried a cost the team wanted to bring down.
Decision
Two changes aimed at cost: migrating observability from Datadog to the Grafana stack, and scaling staging-environment services to zero.
Outcome
Observability moved from Datadog to the Grafana stack, and staging services scale to zero. The saving was significant. Done together with the team.
GrafanaDatadogCost optimizationTIER Mobility
Customer-service automation
Problem
Customers contacted customer service through the TIER app about a set of common issues.
Decision
Built on a third-party customer-communication platform (Intercom), connected through an event-driven design with Kafka so the integration stays decoupled, at the cost of eventual consistency and some duplicated data.
Outcome
Users resolve the most common issues directly through a bot in the TIER app, backed by scalable APIs and data pipelines.
IntercomKafkaNestJSDynamoDBTIER Mobility
Smarter field operations
Problem
Field teams using the Ranger App had to decide where to provide service, such as where to move vehicles and which batteries to change.
Decision
Recommendations are driven by operational data analyzed with the data team: where rides start and end, grouped on Uber's H3 hexagonal grid.
Outcome
The Ranger App recommends optimal parking spots and prioritizes battery swap tasks, so Rangers know where to go to move a vehicle or change a battery.
TypeScriptAWSH3Machine learning
Craft
How I work
Boring where it should be boring, careful where it matters.
Working notes and small models from FinOps practice: measure first, optimize with the trade-offs in view, and move cost into the daily workflow.
Observability firstYou can't optimize what you can't attribute. Unify metrics, traces, logs and cost signals so every dollar maps to a service, a team and a request path.
Pragmatic optimizationRightsizing, scheduling, storage tiers, caching and decoupling, each paired with a reliability guardrail so savings never come out of p99.
Shift-left cost cultureCost is an architectural constraint, like security and latency. It is reviewed in pull requests and owned by squads, not discovered on the monthly invoice.
Observability-driven cost pipeline
Correlate what the infrastructure reports (Prometheus metrics, trace spans, container utilization) with what the cloud bills, so waste has an owner and a price.
Illustrative model at $0.04 per vCPU-hour, rounded from the on-demand AWS Fargate rate for Linux/x86 in US East (N. Virginia): $0.000011244 per vCPU-second, about $0.0405 per hour (checked October 2026). Compute only; real rates vary by region, instance family and commitment. Substitute your own rates and measured utilization. The point: headroom for p99 is a cost you choose; the rest of the idle capacity is waste.
Cost visibility at the service and PR level, not the monthly bill level.
Shift-left FinOps framework
Move cost ownership into the developer workflow: pull requests, CI/CD gates and squad dashboards.
Pull request
Infrastructure-as-code changes open a pull request like any other change. Cost review happens where the decision is made, not weeks later.
Cost diff
CI plans the change and prices the delta (tools such as Infracost do this for Terraform), then posts the monthly difference as a PR comment.
Guardrails
Policy-as-code checks budgets and rules: instance families, storage classes, cross-AZ patterns, tag completeness. Hard stops for violations, warnings for judgement calls.
Deploy
Resources ship with ownership labels (service, team, environment), so the next bill can be allocated without guesswork.
Squad dashboard
Each squad sees its own spend next to latency and error rate, with budgets and anomaly alerts, reviewed in the same rituals as reliability.
Estimated monthly cost change (example)
+ new NAT gateway, eu-central-1 + $ 33
+ 2 x m6i.large (service: orders) + $ 140
~ log retention 90d -> 30d - $ 18
Net + $ 155
Budget (squad: orders) within limit
Every squad treats cost as a core quality attribute, next to latency and reliability.
Architectural optimization patterns
A playbook, not a promise: each pattern comes with the trade-off it introduces and the guardrail that keeps reliability intact.
Patterns
Workload packing: set requests and limits from observed usage, then bin-pack.
Spot or preemptible capacity with autoscaling for interruption-tolerant work.
Rightsizing from p95 utilization over a representative window.
Trade-off
Tighter packing and interruptible capacity cut idle spend but raise noisy-neighbor and interruption risk.
Guardrail
Keep latency-critical paths on on-demand capacity, use disruption budgets and graceful drains, and roll changes out behind SLO checks.
What to track
% waste removed with p99 unchanged
Idle CPU and memory share
Interruption impact on error rate
Patterns
Tiered retention: hot, warm and cold by access pattern.
Cold-path compaction and compression for telemetry and logs.
Lifecycle rules and TTLs set per data class.
Trade-off
Cheaper tiers add retrieval latency and fees, and compaction costs compute. Savings only count if the data is rarely read.
Guardrail
Set retention per data class and compliance need, test restore time, and alert on buckets missing a lifecycle rule.
What to track
Cost per GB-month by tier
% of data in the right tier
Retrieval fees as a share of storage cost
Patterns
Minimize cross-AZ traffic with zone-aware routing.
Route service-to-service traffic internally (private endpoints) instead of over public paths.
Cache and compress where the same bytes cross a paid boundary repeatedly.
Trade-off
Zone-local routing saves transfer cost but concentrates load and weakens failover if pushed too far.
Guardrail
Keep multi-AZ for failure domains, prefer zone-aware routing with fallback, and measure traffic before changing topology.
What to track
Cross-AZ GB per request
NAT and egress share of the bill
Cache hit rate on paid paths
Waste goes down while p99 stays flat. That is the only result worth reporting.
Prototype: ask your cloud bill
A proof of concept for AI-for-FinOps: an MCP service using Claude in Azure, connected to Copilot Studio, so Microsoft Teams Copilot can query cloud account costs, usage metrics and optimization opportunities.
Teams Copilot→Copilot Studio→MCP service (Claude, Azure)→Cloud cost & usage data
Spend broken down by account, service or tag over a chosen period, in plain language.
The utilization and consumption metrics behind the spend, so a number comes with its cause.
Candidate savings, such as idle or oversized resources, for an engineer to review. Suggestions, not automatic changes.
Built at Exxeta as a proof of concept, so it is not open source. The value of MCP here: one typed tool interface, so the assistant answers from real cost data instead of guessing.
Bring the tools to management, not management to the tools.
Senior backend engineer designing and building new features, working as a Golang, Kotlin and Python engineer on an Automobility project. Built and maintain the Grafana stack on Amazon EKS to centralize telemetry across teams, with service names kept consistent across logs, metrics and traces. Designed the concept for a FinOps Hub on Azure FinOps tools with a Power BI dashboard, and built a proof of concept for AI-for-FinOps: an MCP service using Claude in Azure, connected to Copilot Studio, so Microsoft Teams Copilot can answer questions about cloud costs, usage and optimization opportunities. Presented the proof of concept to the customer.
KotlinJavaQuarkusGolangPythonGrafanaAmazon EKSMCPClaudeAzureFinOpsPower BI
Aug 2022 – Feb 2024 Remote
Technical Lead · TIER Mobility SE
Planned and designed the product roadmap with the Project Manager in the TierX Swarm team. Built on TypeScript, NestJS, Nx and AWS, embedding event-driven architecture, domain-driven design and Apache Kafka into core solutions. Helped optimize cloud usage by scaling staging-environment services to zero, and helped migrate from Datadog to the Grafana stack, with significant cost savings. Built Ranger App automation for field operations, using ride start and end locations on the Uber H3 grid (with the data team) to decide where Rangers provide service. Automated the customer-service workflow with an Intercom bot backed by scalable APIs and data pipelines.
Built an infrastructure monitoring system with the ELK stack and Beats for the Titan project (now Open FBA). Led PushBox, an internal tool for Kieler Nachrichten to replace Trello with a private database. Built composable components for Open FBA, a no-code platform research project. Replaced error-prone Excel price calculations with a C# desktop application (HKD Calculator).
TypeScriptVue.jsGraphQLElasticsearchMongoDBC#.NET
May 2013 – Oct 2014 Kuala Lumpur, Malaysia
Junior Software Engineer · Exact
Designed and implemented dashboards for the Exact Cloud platform in the Financial+ Scrum team. Also served as President of the Social and Responsible Team, leading CSR and community initiatives across the Asia Development Center.
C#Visual Basic.NETMS SQL
Background
Education and credentials
Education
M.Sc. Information Engineering Kiel University of Applied Sciences · 2017 – 2019 Technology and Systems Development. Thesis: touch interactions in a no-code programming interface using HTML5 Canvas.
B.Sc. Software Engineering Asia Pacific University, Kuala Lumpur · 2010 – 2013 Dual degree with Staffordshire University, UK.
Certifications
FinOps Certified Practitioner The Linux Foundation (FinOps Foundation)
AWS Certified Solutions Architect – Associate Amazon Web Services
Security Professional Asia Pacific University, 2012
Languages
Persian Native
English Advanced (C1)
German Intermediate (B2)
Contact
Let's talk
Open to interesting conversations about distributed systems, cloud cost and engineering teams.