Ahmed Faraj

Ahmed Faraj
Faculty of Computers & Informatics
Suez Canal University
Ismailia, Egypt
ahmed.faraj.cs gmail com

Fri Sep 4 12:52:59 EEST 2026

GitHub GitLab LinkedIn Email Codeforces Kaggle YouTube CV (PDF)

Daily driver: Neovim, tmux, git.

I am a software engineer at Datajar in Berlin, working remotely from Egypt on the data ingestion and attribution pipelines behind an analytics platform: Shopify, Shopware and the ad networks in one side, a tenant-isolated warehouse out the other. I am in my final year of Computer Science at Suez Canal University.

My interests are operating systems, distributed systems, and the gap between what a specification says and what deployed software actually does. Most recently I measured that gap: I built mcpprobe, a harness that installs an arbitrary Model Context Protocol server into a locked-down container and drives it through the protocol, and ran it over the complete frame of 6,106 registered servers. Two in five never complete a handshake. The dominant deviation turned out to belong to three SDKs rather than to thousands of authors, which is now the first deployment evidence attached to SEP-2145.

Research

First page of Does Your MCP Server Actually Follow the Protocol?

Does Your MCP Server Actually Follow the Protocol?, A. Faraj (Preprint, Aug 2026)
- Executed 6,106 registered servers in isolated Docker sandboxes; only 61.1% (95% CI 59.8–62.3) complete a protocol handshake as published.
- Traced an 88.4% shared deviation to defaults in three SDKs rather than thousands of independent author decisions, supplying the first deployment data to open spec proposal SEP-2145.
- Surfaced a silent-execution class: 129 servers run tools on arguments their own declared schema rejects, violating an explicit MUST in the specification.
- Filed 15 responsible-disclosure reports for servers that crash or hang on a malformed frame; several are already fixed.
Harness & dataset SEP-2145 SEP-1303

Open Source

CNCF, Cloud native, Jun 2026 – Present

pipecd#6979 (open) Fixed a scheduler cancel race that abandoned a running stage and ran rollback in parallel with it. reported as #6978
pipecd#6981 (merged) Fixed group-mention lookup so failed-deployment Slack alerts reach the right on-call group.
kubestellar#3870 (diagnosis) Root-caused ServiceAccount secrets never reaching edge clusters to a metadata-only apply path in library-go, on someone else’s bug report.
karmada#7721 (open) Added MPIJob dependency interpretation so mounted ConfigMaps and Secrets follow the job to member clusters. reported as #7720
karmada#7733 (open) Closed a leaked response body on early returns in the aggregated request handler.
vitess#20364 (merged) Renamed reserved-connection reconnect test packages to describe what they actually exercise.
oras#2087 (merged) Fixed a typo in the log entry separator constant.
kured#1372 (open) Validated the sentinel file path in NewFileRebootChecker instead of failing later at reboot time.

GitLab, Development Kit, main repo, Oct 2025 – Jan 2026

gdk!5345 (merged) Resolved false port conflicts between Nginx and Zoekt that blocked local environments from starting.
gdk!5291 (merged) Added TTY guards before destructive backup and log operations.
gdk!5298 (merged) Fixed the TTY tests that were blocking CI.
gitlab!207789 (merged) Rewrote the API pages curl examples into the multi-line form the docs style guide asks for.
gitlab!207557 (merged) Added descriptive alt text to the ChatOps deployment images.
Hackathon (2nd globally) Placed 2nd worldwide in the GitLab Hackathon among 100+ contributors, Oct 2025.

Elsewhere, Rust, Go, 2025 – Present

solang#1988 (merged) Lowered extendTtl and extendInstanceTtl in the Soroban codegen path of the Solidity compiler.

Employment

Datajar, Software Engineer, Nov 2025 – Present
AI Data Analytics · Berlin, Germany (remote)

Designed the ingestion path across three services, from the source APIs to the warehouse.
Wrote the integrations for Shopify, Shopware and the Meta and Google ad APIs, and took the pipeline to ten thousand requests a minute.
Built the attribution path for PostHog, Google Analytics, GTM and the Meta Pixel.
Enforced tenant isolation with a MinIO registry behind a WorkOS auth layer.

Software

Orbit Onboard on AWS Python · AWS Lambda · SQS · DynamoDB Jun 2026	A new contributor opens an unfamiliar repository and has no idea what matters. Orbit Onboard reads the GitLab Knowledge Graph, ranks 13K+ definitions by call centrality, and writes the orientation a maintainer would have written by hand. This is the architecture that runs it as a multi-user service. Knowledge Graph agent ranks 13K+ symbol definitions by call centrality to produce a reading order for a codebase. Serverless AWS backend: API Gateway with Cognito JWT, DynamoDB cache, SQS queue into a container worker Lambda, S3 + SNS out, DLQ for failures. Under 500 ms on a cache hit, zero idle compute, WAF and KMS on the edge. Built for the GitLab Transcend Hackathon; submitted as the Manara AWS SAA graduation project. Demo Architecture Source
mcpprobe Python · Docker · JSON-RPC Jul 2026	The measurement harness behind the conformance study. Installs an arbitrary MCP server, runs it under --cap-drop ALL with no host mounts, and grades it against the protocol version it negotiates rather than the latest spec. Eight conformance checks over the protocol lifecycle: handshake, ping, tool schema meta-validation, unknown-tool and invalid-argument handling, malformed-frame survival, stdout purity. Two execution modes (ephemeral online, and two-phase offline where measurement runs with --network=none), verified to produce statistically indistinguishable results. Every reported number regenerates from released raw transcripts at a pinned harness tag. Repository
Star Pupil TypeScript · React Flow · LLM Jul 2026	Teaching something is the fastest way to find out you never understood it. Star Pupil gives you a student that knows nothing, believes exactly what you tell it, and sits an exam you never see beforehand. Every mark it loses traces back to a specific sentence in your lesson. An examiner seals a concept map with prerequisite edges before the lesson starts; the learner only ever sees the reachable frontier, so the exam is not a checklist walk. Misconceptions are grounded in quoted source text and propagate along the graph, so a lost mark traces to the root sentence rather than the nearest mistake. The report card splits coverage from accuracy: a confident 6/6 on half the subject is a different result from a shaky 6/6 on all of it. Repository
Stet TypeScript · Next.js · LLM Jul 2026	Re-reading your notes feels like studying and is not. Recognition is cheap and it fakes competence. Stet plants misprints in your own notes, sets the page as a galley proof, and hands you the red pencil. Three-stage generator: edits applied by exact string match in code, then a fact desk that must prove the edit false, then a copy desk forbidden from domain knowledge that kills anything with a stylistic tell. Scores the false strike (correct text you struck anyway) because knowledge you will abandon under pressure is the thing no quiz app measures. Code mode plants bugs that still compile; in testing it turned a Fisher–Yates shuffle into Sattolo’s algorithm. Repository
Event Streaming System Go · Kafka · Redis Ongoing	A distributed event pipeline that records and processes real-time events with a top-N popularity ranking, exposed over REST. Producer/consumer pipeline over Kafka with Redis-backed top-N ranking under concurrent writes. REST endpoints for detail lookup and querying, with explicit concurrency and error handling. Walkthrough
jlox Java · JVM 2025	A recursive-descent interpreter for the Lox language on the JVM: scanner, parser, resolver, tree-walking evaluator, closures and lexical scoping, with AST classes generated by the build. Post-order AST evaluation with a separate resolution pass for lexical scoping and closures. AST class generation and unit tests driven from a Makefile; Java FFI for native interop. Repository
Jack OS Jack · HDL · VM 2025	An operating system for the 16-bit Hack computer, written in Jack: memory allocator, math library, screen and keyboard drivers, string and array types, system I/O. Built bottom-up from NAND gates through a stack VM to the OS layer, as the capstone of Nand2Tetris. Repository

Education

Suez Canal University, B.Sc. Computer Science (Sep 2023 – Jul 2027)
Faculty of Computers & Informatics · Ismailia, Egypt
GPA 3.65 / 4.00, ranked 3rd in class, Academic Excellence Award.
Coursework: data structures, algorithms, operating systems, databases, computer networks, distributed systems.
Transcript

Awards

Competitive programming

ACPC Finalist: 84th at the Arab Collegiate Programming Contest, 59th at ECPC, 1st in university. ICPC
Meta Hacker Cup: advanced to Round 2, 1,480 of 5,972. Meta
Codeforces Expert, top 1% in Egypt. Codeforces
LeetCode Knight, top 2% globally. LeetCode
3,000+ algorithmic problems and 100+ SQL problems solved. AtCoder

Certifications All certifications

AWS Certified Solutions Architect, Associate (SAA-C03). Credly
Redis Certified Associate Developer. Redis

Teaching and Service

YouTube: technical videos on AI agents, systems programming, and competitive programming. Channel
ICPC SCU, IT Head: mentored 1K+ students in data structures and algorithms, authored 50+ ECPC problems. Archive
GDG SCU, IT Specialist: ran programming fundamentals sessions for 100+ students. Archive
CP-Template: competitive programming templates and references used by the SCU team. 13 stars