Cloud Platform · Backend · Applied AI

Amirhossein
Faghihifar

I build the cloud foundations and AI systems that products run on.

Founding platform engineer at SEPANTA. Nine years across Google Cloud, Kubernetes, Terraform, FastAPI and NLP — from Tehran, to the UK, to three markets in Southeast Asia.

Explore the work Get in touch
Based in United Kingdom Now SEPANTA (ex-Goodfolio)
base production cell
0years engineering
0calls processed by SumX
0transactions / day, 99%+ uptime
0CI-only Terraform pipelines
0countries, one platform
status: building sepanta-org / platform

Currently

Building SEPANTA's entire Google Cloud organisation, solo and greenfield: Shared VPC networking, private GKE Autopilot cells, org-policy and IAM governance, and keyless CI through Workload Identity Federation. Every infrastructure change ships through review-gated Terraform pipelines on GitHub Actions. No service-account keys exist anywhere, and nothing is ever applied from a laptop.

  • ✓ shared VPC · private GKE Autopilot · Cloud Run
  • ✓ 11 review-gated, CI-only Terraform pipelines
  • ✓ org-wide observability, audit & cost visibility
  • ◌ reliability & incident response across products
01 / experience

Deploy log

Nine years, five companies, one through-line: taking AI and data systems from idea to something reliable in production.

  1. Sep 2024 — present
    SEPANTA formerly Goodfolio · United Kingdom
    Cloud Platform & DevOps Engineer (founding) Feb 2026 →
    • Built a greenfield Google Cloud organisation solo: Shared VPC, private GKE Autopilot, org-policy and IAM governance, keyless CI via Workload Identity Federation.
    • Automated all infrastructure delivery through 11 review-gated, CI-only Terraform pipelines; set up org-wide observability and cloud-cost visibility.
    • Own reliability across products: Cloud Run and GKE deployments, monitoring, logging, secrets, incident response.
    Software & Machine Learning Engineer Sep 2024 → Feb 2026
    • Designed FastAPI services over PostgreSQL and Neo4j, with CI/CD to GCP.
    • Shipped generative-AI product features, vector search on AlloyDB, and self-hosted LLM observability (Langfuse on GKE via Terraform).
    GCPGKE AutopilotTerraformFastAPINeo4jLLMs
  2. Sep 2020 — Sep 2024
    Mofid Securities Iran's largest brokerage · Tehran
    Software Technical Lead Mar 2023 → Sep 2024
    • Led and mentored a team of four, owning architecture and technical direction.
    • Architected a RabbitMQ microservice pipeline handling 20,000+ transactions a day at 99%+ uptime, monitored with Prometheus, Grafana and Loki.
    AI & Software Engineer Sep 2020 → Mar 2023
    • Deployed AI systems firm-wide and led SumX, a speech-analytics platform spanning the company's communication channels.
    • Built sentiment and intent models (>85% F1) and RBAC/search features with Flask and FastAPI.
    PythonGoRabbitMQDocker SwarmASR / NLPMongoDB
  3. Aug 2019 — Oct 2020
    Vira Afzar Adan Faraazin · Tehran
    Machine Learning Engineer (NLP)
    • Trained English–Persian neural machine translation (BLEU 45, MarianNMT) over 200M+ tokens; served it in production at 100,000 words per minute.
    MarianNMTseq2seqFlaskDocker
  4. 2017 — 2018
    Raypo · ParsaTech Tehran
    Backend Developer · Server & Support Engineer
    • Customer-loyalty platform with Django, DRF, Celery and Redis. Deployed and supported IPTV servers on Linux.
    DjangoCeleryLinux
02 / projects

Selected systems

Each of these is a pipeline of some kind. Hover a card to spin it up.

SEPANTAcloud platformbackendapplied AI

NuStream formerly Libera

Retail data-collection and insight platform across Indonesia, Pakistan and the Philippines. Field photos of shelves and receipts become shelf-level insight for brands.

  • Own the Google Cloud foundation: a cell-per-country architecture, each country its own private GKE Autopilot cluster plus co-located data plane, so data stays in-region and cells scale and fail independently.
  • Built the admin, Stores, Products and Submissions APIs on Cloud Run, plus their GitHub Actions CI/CD.
  • Replaced manual product tagging with an LLM auto-categoriser: a consistent three-level taxonomy (~130 leaf categories) across tens of thousands of products, written into the Neo4j graph.
GKE AutopilotShared VPCTerraformCloud RunNeo4jQdrantLLMs
Read the case study →
SEPANTA / Finspector.AIcloud & devopsagentic AI

Finspector

AI-driven compliance for financial marketing. Inspects promotions across files, social channels and websites against FCA or custom rules, flags issues, keeps an audit trail.

  • Keyless CI/CD on GitHub Actions via Workload Identity Federation, with dependency and container vulnerability scanning gating every deploy.
  • Designed the public edge: Cloud Load Balancing, domain management and Cloud Armor; centralised IAM.
  • Made the agentic system observable and defensible: Langfuse LLM tracing on GKE, logging, tracing and alerting.
  • Developer on the social-media processor agent; built the shared Valkey cache layer.
Cloud RunPub/SubGoogle ADKCloud ArmorWIFLangfuseValkey
Read the case study →
Mofid Securitiesarchitecture leadspeech · NLP

SumX

Turns call-centre audio into searchable, analysable data. 10k+ calls and 300+ hours of audio a day; over 5 million calls across the project's life.

  • Led the end-to-end architecture: VAD, ASR, intent and sentiment on every dual-channel call; derived features such as silence %, talk speed, agent/customer talk share and forbidden words.
  • Built the multi-filter search dashboard, a QA workspace and multi-level management reporting with specialised RBAC.
  • Stood up the runtime myself: a 12-node Docker Swarm cluster running 20+ services over RabbitMQ and MongoDB, with Prometheus, Grafana, Loki, Alertmanager and Metabase.
PythonASR / VADRabbitMQDocker SwarmMongoDBGrafana
Read the case study →
Mofid Securitiesdistributed systemsVoIP · SIPLLM

Smart Outbound Call Service

Automated, interactive outbound voice. Calls a customer, speaks a personalised message, listens, and lets an LLM choose the next branch. No agent on the line.

  • Built the stateful FastAPI controller end to end: call flows, customer lists, per-call progress, orchestration of every call.
  • Designed the WebSocket action protocol (initiate, play, listen, end) that let the Go SIP/RTP media service stay completely stateless and disposable.
  • Integrated TTS, ASR and an LLM so flows adapt to what the customer actually says. Hundreds of calls an hour, tens of agent-hours saved per campaign.
FastAPIGoWebSocketsSIP / RTPTTS · ASRLLMs
Read the case study →

also Faraazin MT platform — English–Persian neural machine translation trained on 200M+ tokens and served in production at 100k words/min. · MSc thesis — end-to-end speech-to-text translation that beat a cascaded baseline by 5.7 BLEU, on two speech corpora I built with a Telegram crowdsourcing platform.

03 / skills

terraform plan

Everything I'd bring to your stack, declared as code.

amir@platform — zsh

  


  

    
Cloud & platform
Google CloudGKE / AutopilotCloud RunTerraformShared VPCWorkload Identity FederationCloud ArmorHelmAnsible

    
Backend & data
PythonGoFastAPIDjangoGinPostgreSQLAlloyDBNeo4jMongoDBRedis / ValkeyRabbitMQPub/Sub

    
AI / ML & NLP
LLMs & agentsvector search (Qdrant)MarianNMTseq2seqASR / TTSsentiment · intentscikit-learnPandas

    
Delivery & observability
GitHub ActionsDocker / SwarmLangfusePrometheusGrafanaLokiMetabase

  







  

    04 / deploy
    
Let's build something that stays up.

    
Open to senior platform, backend and applied-AI roles, and to interesting conversations about any of the above.

    

      a.faghihifar@gmail.com
      LinkedIn ↗
      GitHub ↗