Sergo Azizbekyan

Hi, I'm

Sergo Azizbekyan

Software & AI Engineer

sergo.azizbekyan@gmail.com · sergo.azizbekyan.online · github.com/Pro100Sergosha · linkedin.com/in/sergo-azizbekyan-8341a8307

Backend engineer with 3+ years in production. I build at Tarsai Systems across Python backends, AWS infrastructure and LLM integrations.

Download CV GitHub LinkedIn

About

I build backends that keep working when the things around them fail: third-party integrations isolated from core logic, retries and dead-letter queues, observability from the first deploy. I have taken a product from zero to production alone and led delivery in small teams. I work AI-first — agents write much of the code, while the design, the reviews and the trade-offs stay with me.

Experience

  1. Software & AI Engineer · Tarsai Systems

    Founding team of an early-stage AI-native software company: client systems, AI integration, and a venture lab that takes new products from idea to production in weeks, with AI agents writing code and engineers owning every decision. First lab product: Binadar.ge, which aggregates real-estate listings from myhome.ge, ss.ge and livo.ge and alerts agents when one matches a client's brief. Serverless AWS: Lambda, SQS, EventBridge, DynamoDB, Cognito, CDK; Astro + React frontend.

    • Split the monolithic API Lambda into three functions by trust boundary, so a bug in a public handler no longer runs with the widest IAM rights in the system (account deletion, full-table writes).
    • Made manual filter runs asynchronous (202 Accepted + background worker) with per-caller rate limiting, lifting the 29-second API Gateway ceiling that capped manual runs below scheduled ones.
    • Designed a PaymentProvider port and a Keepz adapter built from the public docs before credentials existed; payment callbacks are treated as hints and every plan change is re-verified against the provider.
    • Built the account-wide match inbox on an existing index with no migration, moved state filtering server-side (fixing saved matches silently vanishing past the first page), and added numbered pagination in the frontend.
    • Restructured the monorepo into a shared core + per-service layout across 8 services, unblocked production deploys broken by an AWS Lambda/SQS concurrency change, and added a human approval gate before production.

  2. Python Backend Developer · Making Science

    • Architected a 9-service FastAPI backend behind APISIX with 5-tier rate limiting, handling up to 100 req/s. A shared resilience library (circuit breaker, retry, bulkhead, DLQ) gives every service the same failure handling instead of each one reinventing it.
    • Built one integration layer for 3 external providers — Keepz payments, Didit identity verification, Google OAuth — so an outage at any of them degrades a single feature, not the core flow.
    • Cut deploys from ~30 to ~5 minutes: GitHub Actions detects which services changed and rebuilds only those.
    • Set up Grafana + Loki across all 9 services, so logs from every service are searched in one place during an incident.
    • Led delivery in a 2-person team: reviewed and merged 58+ PRs, added RabbitMQ for async messaging and MinIO for file storage.

  3. Backend Developer · Dealin

    • Took the product from zero to production as the sole developer: 25+ endpoints serving up to 1,000 users on Django 6, DRF, PostgreSQL, Celery and Redis.
    • Shipped the full auth flow from day one: email verification, 2FA, password reset, JWT and Google OAuth.
    • Moved email delivery out of the request cycle into a Celery + Redis queue, so a slow mail server no longer holds up API responses.
    • Every push runs ruff, tests and a Docker build in GitHub Actions; the schema evolved through 16 migrations with index tuning along the way.

Projects

LangChain Text Analyzer

Classifies topics, scores sentiment and answers questions about a text, using retrieval over Milvus and Google Gemini.

LangChain · Gemini · Milvus · FastAPI · SQLModel · Docker

FastAPI WebScraper

Scrapes JavaScript-heavy pages at ~200 pages per 2 minutes and turns them into structured Markdown ready for RAG and LLM pipelines, on a cron schedule.

FastAPI · nodriver · asyncio

Fitness Tracker API

REST API for workouts — exercises, plans, sessions and weight logs — fully covered with pytest.

Django · DRF · PostgreSQL · Docker · pytest

Open Source

herdr-profiles PR #4 · open

Added per-profile environment variables to Herdr's Chrome-style profiles, so each profile runs tools under its own account — for example Claude Code on a subscription in one and Vertex AI in another. Along the way, stopped a broken profiles.json from being silently overwritten, and added editing from the popup and Windows Terminal support under WSL.

Rust · +405 / −39 · 36 tests · clippy clean

Skills

Backend
Python, FastAPI, Django, DRF, SQLAlchemy, SQLModel, Alembic, Celery, asyncio
Data & Messaging
PostgreSQL, Redis, RabbitMQ, Milvus, MinIO
AI
RAG, LangChain, Google Gemini, Anthropic Claude API, LLM APIs, Embeddings
DevOps & Cloud
Docker, GCP, AWS (EC2, AuroraDB, S3, Serverless, ElastiCache, Bedrock), GitHub Actions, CI/CD, APISIX, Nginx, Linux, Grafana, Loki
Tooling
pytest, ruff, uv, Git, Swagger, ReDoc, nodriver, BeautifulSoup4, Claude Code, Copilot, Codex

Education & Certifications

Let's talk