Open to new roles · available now

Israel · Israeli citizen · full EU working-hours overlap (UTC+2/3)

Beniamin Levin Hands-on Engineering Team Lead / Tech Lead — still coding ~80% of the time.

For 9 years as Lead Software Engineer at Netcracker I managed up to 11 direct reports in Europe and India while coding ~80% of the time in Java and Spring Boot. I owned technical design and delivery of the team’s 13 telecom OSS/BSS microservices — ~3,500 events a second, Kafka on Kubernetes in AWS, GCP, and Azure — for operators such as Vodafone, Telenet, and Rakuten.

I also owned the roadmap, hiring, performance reviews, and mentoring, and brought AI into the team’s daily work: Cursor for development, and a triage agent that took defect triage from ~2 hours to ~30 minutes. On my own time I build FixedYet, an AI agent that checks GitHub bug reports by running code, and run ikite.fyi, a production Go service.

Java · Spring Boot 3 · Kafka · Kubernetes · AWS / GCP / Azure · Python · Go · AI-assisted engineering

Direct reports
up to 11
Europe and India, while coding
Engineers hired / later promoted
9 / 5
4 to Technical Lead, 1 to Senior
Incoming defects
−75%
Order Capture, over two years
Average triage time
~2h → ~30min
AI agent adopted by 40 teams
Beniamin Levin — Engineering Team Lead / Tech Lead

Selected Engineering Impact

Six engineering outcomes I led or contributed to at Netcracker — from production performance and defect reduction to team growth and AI-assisted workflows — each with its scope and my role. I am happy to walk through how each was measured.

  • ~4s → ~0.5s

    Order-entry response times, 30–88% lower

    Across critical order-entry workflows, by profiling with JFR, removing duplicate service calls, adding caching, and tuning Oracle SQL in the Java/Spring backend.

  • 9 / 5

    Engineers hired / later promoted

    Hired 9 engineers in India, including a 5-engineer CPQ pod in Pune (70+ interviews in three months), and onboarded them on site during two multi-month handovers. Four were later promoted to Technical Lead and one to Senior Software Engineer.

  • ~40 → ~10

    Incoming Order Capture defects a month

    Down ~75% over two years, by fixing recurring defects at their root cause and tightening code reviews and CI/CD validation.

  • ~2h → ~30min

    Average defect-triage time

    I wrote the agent skill behind Netcracker’s Order Capture triage agent, grounded through MCP tools (Jira, GitLab, Jenkins) and RAG over docs and past tickets. Adopted by 40 teams for 1,000+ tickets a month; an estimated ~1,500 engineer-hours saved a month.

  • ~3,500/s

    Events through the team’s 13 microservices

    Order entry, CPQ, product catalog, and customer information management under a 99.99% availability SLA. Java/Spring, Quarkus, Kafka, and Kubernetes across AWS, GCP, and Azure.

  • −40%

    Unit-test creation time

    After I rolled out Cursor for AI-assisted development in the team’s Java/Spring and Angular/React code, with repository-aware workflows for coding, refactoring, code analysis, and test generation.

Professional Experience

Telecom OSS/BSS, automotive, and mobile platforms — leading teams while staying in the code.

Netcracker Technology

Jul 2017 – Sep 2026

Lead Software Engineer

Ra’anana, Israel

Team lead for telecom OSS/BSS order entry, CPQ & product catalog

Java · Spring Boot · Quarkus · Kafka · Kubernetes · AWS / GCP / Azure · Angular
  • Managed up to 11 direct reports in Europe and India while coding ~80% of the time in Java/Spring. Owned the team’s roadmap, turned new product requirements into technical designs, and ran sprint and release planning, performance reviews, mentoring, code reviews, and incident resolution. Delivered releases to 5–6 operators a year, including end-to-end order entry for Telenet’s BSS transformation (mobile, TV, fixed; residential and business) and the Rakuten and Vodafone Hungary launches.
  • Hired 9 engineers in India, including a 5-engineer CPQ pod in Pune (70+ interviews in three months); onboarded them on site during two multi-month handovers. Four were later promoted to Technical Lead and one to Senior Software Engineer.
  • Owned technical design and delivery of the team’s 13 microservices for telecom OSS/BSS (order entry, CPQ, product catalog, customer information management) on Java/Spring, Quarkus, Node.js, Kafka/RabbitMQ, PostgreSQL/Oracle, Cassandra, Redis, and Elasticsearch on Kubernetes (AWS, GCP, Azure), handling ~3,500 events a second under a 99.99% availability SLA.
  • Coded REST APIs for sales orders, product characteristics, and customers, and Angular order-entry UI features and fixes.
  • Diagnosed and resolved production JVM issues: a classloading deadlock in TELUS’s CPQ quotation engine and full-GC pauses on Change Ownership screens; also backported the Kubernetes 1.26 HPA autoscaling configuration to CPQ 22.3.
  • Decreased incoming Order Capture defects ~75% over two years (from ~40 to ~10 a month) by fixing recurring defects at their root cause and tightening code reviews and CI/CD validation; maintained the team’s Jenkins and GitLab CI pipelines.
  • Shortened average defect-triage time from ~2 hours to ~30 minutes by writing the agent skill behind Netcracker’s Order Capture triage agent on Claude, Kimi, and GPT, grounded through MCP tools (Jira, GitLab, Jenkins) and RAG over docs and past tickets. Adopted by 40 teams for 1,000+ tickets a month (an estimated ~1,500 engineer-hours saved a month).
  • Resolved ~2,000 L3 tickets for 30+ operators (Telefónica/O2 UK, TELUS, Nuuday, Charter) typically in 1–3 days, prioritizing escalations with product managers and using Python scripts to parse service logs and query Oracle and PostgreSQL data.
  • Lowered unit-test creation time ~40% by rolling out Cursor for AI-assisted development in the team’s Java/Spring and Angular/React code, with repository-aware workflows for coding, refactoring, code analysis, and test generation.

Netcracker Technology

Jul 2015 – Jul 2017

Senior Software Engineer

Nizhny Novgorod · promoted to Lead in 2017

Order-entry performance

Java · Spring · Oracle SQL · JFR
  • Reduced order-entry response times 30–88% across critical workflows (key responses from ~4s to ~0.5s) by profiling with JFR, removing duplicate service calls, adding caching, and tuning Oracle SQL in the Java/Spring backend.

Symphony Teleca (now Harman)

Feb 2011 – Jul 2015

Principal Software Engineer

Nizhny Novgorod

Connected car, mobile platforms & core messaging

Java · Android
  • Engineered the Java backend and Android client of SiriusXM’s connected-car infotainment platform, shown live at CES.
  • Optimized core messaging infrastructure for flagship Samsung Android devices on site at Samsung HQ in Suwon, South Korea, and built Android UI and widgets for Nokia and Savelli.

Onlyoffice (then TeamLab)

Nov 2009 – Jan 2011

Software Engineer

Nizhny Novgorod

Enterprise collaboration & document management

C# · ASP.NET
  • Built C# and ASP.NET modules for an enterprise collaboration and document-management platform.

MERA NN

Nov 2006 – Nov 2009

Software Engineer

Nizhny Novgorod

Telecom systems management

Java · High Availability
  • Built high-availability Java web applications for remote administration and monitoring of Nortel communication systems.

Projects & Freelance

Systems I build and run on my own — designed, written, deployed, and operated by me. Every featured project has public source code, if you would rather read current code than take my word for it.

ikite.fyi

Feb 2023 – Present

Case study · sole engineer & operator

Live at ikite.fyi

Go source on GitHub

Production wind-monitoring app for kitesurfers on the Israeli coast

Go · MySQL · Oracle Cloud (OCI) · systemd · Docker · Telegram

Kitesurfers on the Israeli coast decide whether to drive an hour based on wind that changes in twenty minutes. Public sources are scattered, slow, and mostly Hebrew PDFs. I built the thing I wanted to exist, then kept it running.

Architecture

Windguru + Israeli stations → Go collectors every minute (systemd timers) → MySQL → Go web app → charts, beach cameras, Telegram and browser alerts

Live station readings, historical charts, beach cameras, and wind, rain, and storm alerts, on Oracle Cloud. Go collectors ingest Windguru and Israeli station data into MySQL every minute and send wind and storm alerts through Telegram and browser notifications. They run on independent systemd timers, so one flaky upstream cannot stall the others; development runs on a Docker-based setup.

Live rain and storm alerts

RainViewer radar (every 10 min) + EUMETSAT MTG satellite lightning → rain-cell motion between consecutive frames → projected forward → Telegram alert (@KYstormbot) when two frames agree rain arrives within 60 min

I added live rain and storm alerts from radar and satellite data: ikite tracks rain-cell motion across consecutive radar images, projects it forward, and alerts through Telegram before rain reaches the beach. Replayed over a month against a rain gauge ~2 km away, it warned of 7 of 10 rain spells a median of ~40 minutes ahead, and it self-tunes after each storm.

Why I rewrote it from PHP to Go

The 2023 PHP version worked, but every collector was a separate cron-invoked process paying full interpreter startup, and concurrent polling meant either blocking or process sprawl. Go gave me goroutine-per-collector concurrency, a single static binary to deploy, and a memory profile that fits comfortably inside a free-tier VM. The 2026 rewrite is the version running today.

What running it taught me

Upstream feeds fail in undignified ways — a station returns yesterday’s reading with today’s timestamp, a PDF layout shifts, a file arrives truncated. Most of the engineering is not the happy path; it is staleness detection, per-source backoff, and alerting that distinguishes “no wind” from “no data”.

FixedYet

Sep 2026 – Present

Solo build · Nebius x NVIDIA Global AI Hackathon entry

Live demo

Python source on GitLab

Demo video (2:54)

Hackathon entry on Devpost

An autonomous AI agent that validates GitHub bug reports

Python 3.12 · asyncio · FastAPI · Pydantic · Docker · MCP server · NVIDIA Nemotron

A bug report shouldn’t be marked fixed because an LLM thinks the patch looks right. FixedYet writes a reproduction test, runs it in a sandbox at the reported version and on current main, and lets deterministic rules, not the model, decide the verdict. It can then test a proposed fix or backport in the sandbox and open a draft pull request. I built it solo, to answer the two questions a maintainer actually has: is the bug real at the reporter’s version, and is it already fixed?

How it works

GitHub issue → Nemotron writes a pytest reproduction → sandbox runs it at the reported version and on main → deterministic verdict → fix or backport, proven in the sandbox → draft PR

NVIDIA Nemotron on Nebius Token Factory writes the pytest reproductions, Docker isolates each run, and deterministic rules, not the model, decide the verdict: fixed, open, or not reproduced. For a bug that is still open it writes a fix; for one already fixed on main it verifies the backport to the reporter’s release line. A change counts only when the reproduction fails before it, passes after it, and the project’s own suite shows no new failure.

Results
  • 11–13 of 15 issues of a SWE-bench Verified subset reproduced per run, at ~$0.03 of model cost each (a small subset, not comparable to the public leaderboards)
  • 8 of 10 correct verdicts on recent live GitHub issues, a separate set run as a check against training-data overlap; the other 2 were flagged at low confidence
  • Four upstream PRs that its fixes and backports led to, including Black (42k stars), Pydantic, and RDFLib (merged); I finished them after maintainer review
  • 1,700+ pytest tests in the project itself
What building it taught me

Passing tests isn’t the same as correct. FixedYet’s first RDFLib fix passed the whole suite but also changed SPARQL MIN and MAX; review caught it, and the merged fix took a different approach. Scope, plainly: pure-Python projects tested with pytest.

Checkout Sentinel

Oct 2026 – Present

Solo build · PayPal AI Hackathon 2026 entry

Python source on GitLab

Demo video (2:42)

Hackathon entry on Devpost

Payment-page skimming detection with verified security-policy remediation

Python 3.12 · FastAPI · Playwright + CDP · PostgreSQL · Docker Compose · NVIDIA Nemotron · AWS EC2 & RDS

One compromised third-party script on a checkout page can skim every field a buyer types, or draw a fake PayPal login over the real checkout. I built Checkout Sentinel to detect that kind of payment-page tampering on PayPal merchants’ checkouts and to verify the security policies that stop it. It produces PCI DSS v4.0.1-aligned evidence for requirements 6.4.3 and 11.6.1; it is not a compliance certification. The AI investigates, deterministic checks verify, and a human approves.

How it works

Headless-browser crawl of the payment page → script inventory → deterministic detectors → AI investigation with read-only tools → proposed CSP + SRI → attack again with the policy enforced → human approval

Playwright with raw CDP inventories the payment page’s scripts, records listener and network activity, and drives a PayPal sandbox checkout up to the approval step. Deterministic detectors flag injected or tampered scripts, unauthorised listeners on payment fields, data sent to unknown destinations and fake overlays. The AI explains each finding and proposes a Content-Security-Policy and Subresource Integrity pins, as data, never code. A proposal passes only when the page, attacked again with the policy enforced, leaks no canary to the attacker, keeps its essential scripts working, leaves the clean control page unchanged and still reaches PayPal approval. No policy is deployed to the shop without human approval, and a post-deploy check rolls it back if the shop does not serve the approved policy.

Results
  • Detection: every rule at precision 1.0, recall 1.0 and false-positive rate 0.0 on the four demo scenarios, with zero findings on the clean control
  • AI investigation: 9/9 VERIFIED and correctly classified (3 scenarios × 3 runs), total cost $0.1255
  • Hosted on AWS (EC2, RDS) since 30 Oct 2026; CI runs Semgrep, pip-audit, gitleaks and an OWASP ZAP baseline scan
Limits, stated plainly

Missing or unobserved evidence fails verification. Detection results come from recorded crawls of a synthetic demo shop; the AI-eval numbers come from live-recorded runs replayed offline (no new live LLM calls). Checkout is verified up to PayPal sandbox approval; a completed capture is not claimed. Each crawl is one browser visit, a sample rather than every shopper’s experience, and in-page listener hooks can be evaded by determined attacker code.

Additional engineering work

  • Sure Communications — Part-Time Freelance Software Engineer 2019 – Present

    Part-time, evenings and weekends, outside full-time employment and in a non-competing domain. PHP/MySQL/JavaScript applications, customer portals, and CRM integrations for UK business clients, including Morgan Motor Company’s dealer sales platform; also built an OpenAI email-reply generator and SurAI, a natural-language-to-SQL BI copilot with read-only guardrails and Chart.js output.

  • Business FM — Full-Stack Architect & Developer 2012 – 2015

    Sole developer of a bespoke PHP/MySQL CRM; migrated it from Azure to AWS EC2/RDS.

Core Competencies & Technical Skills

Core technologies and engineering practices, grouped by discipline. Highlighted skills are my current hands-on work; the rest is the wider stack I have shipped production systems with.

Leadership, Architecture & Delivery

Engineering Leadership People Management Hiring & Interviewing Performance Reviews Team Mentoring Code Review System Design Distributed Systems Microservices Architecture Event-Driven Architecture OOP Design Patterns SOLID Principles L3 Production Support Performance Tuning Release Management Agile / Scrum Kanban Sprint Planning Jira Story Mapping Root-Cause Analysis

Backend

Java Java EE Spring Boot 3 Spring Framework 6 Spring MVC Spring Data JPA Spring Security Spring Cloud (Gateway, Config) Hibernate jOOQ Tomcat JUnit Mockito Maven Gradle Quarkus Multithreading & Concurrency Microservices Go (Golang) Node.js Python

APIs & Security

REST OpenAPI / Swagger gRPC GraphQL OAuth OpenID Connect JWT Keycloak RBAC

Messaging & Data

Kafka RabbitMQ Redis Elasticsearch NoSQL Cassandra SQL PL/SQL Stored Procedures PostgreSQL Oracle MySQL MongoDB

Cloud & DevOps

AWS GCP Azure Oracle Cloud Kubernetes Docker Helm OpenShift Terraform CI/CD Git Jenkins GitLab Argo CD Observability Prometheus Grafana JMX JFR Linux Bash systemd Nginx

AI-Assisted Engineering

Claude Code Cursor LLM Prompt Engineering Agent Skill Design RAG-Grounded Agents MCP Agent Tooling Repository-Aware Code Analysis AI-Assisted Defect Triage LLMs (Claude, GPT, Gemini, Kimi)

Frontend

JavaScript TypeScript React Angular RxJS HTML5 CSS3

Earlier

WebLogic GWT C# ASP.NET PHP Laravel WordPress WooCommerce SOAP XML jQuery Bootstrap Android

Education & Languages

N. I. Lobachevsky State University of Nizhny Novgorod

Sep 2003 – Jul 2008

With Honors

Master’s in Applied Informatics (Computer Science)

Faculty of Computational Mathematics and Cybernetics

Languages

  • English

    Fluent

  • Hebrew

    Professional working proficiency

  • Russian

    Native

Open to new roles · available now

Let’s Connect

I am available now — my Netcracker role wrapped up in September 2026. I am most effective as a hands-on team lead or tech lead: running the team, the roadmap, and the hiring while still writing a good share of the code. Happy to talk about distributed system design, Kafka and Kubernetes in production, growing a team, or getting AI agents usefully into an engineering team’s workflow.

✓ Email copied to clipboard