Skip to content
YA
00Portfolio · v2026

Reasoning systems, hardened networks.

I build systems that reason and systems that hold up under load — machine-learning pipelines for encrypted traffic, autonomous agents wired into live carrier networks, and distributed log platforms that detect anomalies before a human notices them.

View selected workDownload CV
Base
Algeria
Discipline
AI · Security · Networks
Systems
5 documented
Capability areas
20 mapped
Open to research collaboration & engineering rolesyahiaongh

Intelligence — anomaly detection · grounded retrieval · agent guardrails

pointer reactive

01Identity

An engineer who reads the whole diagram — from the packet to the model.

Most problems I take on sit where two disciplines meet: a statistical method that has to survive a real network, or a carrier API that has to survive an adversarial transaction.

I am Yahia Abdeldjalil Benyahia — an AI systems and telecom infrastructure engineer working across detection, anti-fraud decisioning and distributed event pipelines. My work tends to start the same way: read the actual system, find the place where the honest answer and the convenient answer diverge, and build for the honest one.

That has taken me from NumPy implementations of a single neuron, through a self-hosted 5G core, into polyglot pipelines where a Rust detector scores aggregates a Go service just wrote and a Python service explains in plain language. Each step was a different problem; the discipline was the same.

PythonGoRustTypeScriptC++PostgreSQLDockerKubernetes
run sentinelops
git clone https://github.com/yahiaongh/sentinelops.git
cd sentinelops/infra && cp .env.example .env
docker compose up -d
# ~20s to healthy, then:
docker exec -it sentinelops-timescaledb psql -U sentinelops -d sentinelops \
  -c 'SELECT count(*) FROM log_events;'
whoami — zsh
  1. whoami --verbose
  2. Yahia Abdeldjalil Benyahia
  3. AI Systems & Telecom Infrastructure Engineer
  5. FOCUS
  6. Distributed log intelligence · anomaly detection
  7. Telco-aware anti-fraud · agent guardrails
  8. 5G core networks · encrypted traffic analysis
  10. PRINCIPLES
  11. ground every model output in retrievable evidence
  12. make at-least-once delivery safe to replay
  13. measure the system, then trust the measurement
  15. STATUS
  16. open to research collaboration & engineering roles
01

Evidence before output

A model that cannot cite where a number came from is a rumour with good typography.

02

Failure modes are the design

Duplicate delivery, provider outages and non-volatile storage are planned for, not discovered.

03

Measure, then trust

Every real bug in these projects came from reading live behaviour, not from reading green tests.

Trajectory

  1. Foundation

    Core ML implemented from scratch — logistic regression, ReLU, sigmoid — because frameworks hide the derivations that matter.

  2. Networks

    5G core deployed end to end: AMF, SMF and UPF on Free5GC with UERANSIM registration and live PDU sessions.

  3. Carrier APIs

    CAMARA signal semantics and sandbox behaviour mapped while building an anti-fraud guard engine on live carrier data.

  4. Production plumbing

    Distributed log intelligence across Go, Rust and Python — ingestion, detection, grounded querying, observability.

02Engineering DNA

Five disciplines, one habit: find the seam.

The interesting problems live where these disciplines overlap. Each group below names what I actually build, and links to the repository that proves it.

01Intelligence

Models that produce a number someone can check, or a decision someone can defend.

  • Statistical detectionZ-score and EWMA over rolling aggregates — directional where direction is the signal.
  • Retrieval groundingRAG constrained to real rows, with the model's output verified against them.
  • Agent orchestrationPlanner decides which tools run; deterministic layer holds the contract; model layer refines.
  • OptimisationMemetic search, hyper-heuristics and metaheuristics for problems too large to solve exactly.
  • Python
  • NumPy
  • Numba
  • PyTorch
  • TensorFlow
  • Ollama

02Security

Findings that carry their own evidence, and controls that fail closed.

  • Anti-fraud decisioningGrounded verdicts with per-tool provenance and a confidence score that reflects it.
  • Encrypted traffic analysisClassification of QUIC flows from statistical features when payloads are unreadable.
  • Model guardrailsFloor, ceiling and prose-consistency constraints so a model cannot soften a verdict.
  • Threat modellingRate limiting, tenant scoping, fail-closed auth — documented honestly against demo scope.
  • CAMARA
  • Nokia NaC
  • SIM-swap signals
  • Intrusion detection

03Networks

Systems understood at the packet, the protocol and the topology.

  • 5G coreAMF, SMF and UPF deployed end to end; registration and PDU sessions verified.
  • Carrier APIsCAMARA signal semantics, sandbox behaviour and the difference between the two.
  • Edge–cloud placementWhere inference runs, and what that costs in transfer latency against throughput.
  • Protocol workQUIC flow statistics, feature extraction and classifier evaluation.
  • Free5GC
  • UERANSIM
  • Docker
  • Linux
  • QUIC
  • CAMARA

04Systems

Event-driven plumbing that survives duplicate delivery, restarts and partial failure.

  • Event-driven ingestionAt-least-once delivery absorbed with batched, idempotent writes and composite keys.
  • Polyglot servicesGo, Rust and Python in one pipeline, each owning its schema and its metrics.
  • Time-series storageHypertables, continuous aggregates and the poll cadence that ties them together.
  • ObservabilityMetrics scraped, dashboards provisioned in version control, gaps written down.
  • Redpanda
  • TimescaleDB
  • Prometheus
  • Grafana
  • Kubernetes
  • Docker

05Engineering

The unglamorous half: build gates, schema discipline, and CI that blocks bad merges.

  • Self-migrating schemasMigrations embedded at compile time and applied idempotently — no manual psql step.
  • CI as a gateMulti-language lint, test and build jobs required before merge to main.
  • Numerically careful codeOverflow, cancellation and precision handled where the naive version silently fails.
  • Interface engineeringAccessible dialogs, streaming operator consoles, keyboard-navigable data views.
  • Go
  • Rust
  • TypeScript
  • C++
  • React
  • Postgres
  • GitHub Actions

03Selected work

Systems, not screenshots — each one open source.

Two projects carry most of the weight, because they are the ones where the hard part was the system rather than the interface. Every claim here comes from the repository's own documentation, including where it falls short.

Distributed SystemsPython · Go · Rust +7

Distributed log intelligence with statistical anomaly detection

  • Event-driven ingestion over Redpanda into TimescaleDB hypertables
  • Rust anomaly engine: directional z-score on p95/max latency, EWMA on error rate
  • Grounded RAG — model answers verified against real rows, not hallucinated
Languages
Go · Rust · Python · TS
CI jobs
9 required
Tests
Go, Rust, Python, Vitest
License
MIT
Repository
  • Python
  • Go
  • Rust
  • TypeScript
CybersecurityPython · FastAPI · LangGraph +7

Telco-aware anti-fraud verdicts from live CAMARA network signals

  • 9 CAMARA signals via Nokia Network-as-Code, with SDK → REST → sandbox fallback
  • Deterministic rule engine is the contract; the LLM may only intensify risk
  • Live SSE pipeline streamed into a Next.js operator console
Event
GSMA MENA Ignite 2026
CAMARA tools
9
Backend tests
191 passed · 1 skipped
Verdict match
10/10 deterministic
RepositoryOpen live demo
  • Python
  • FastAPI
  • LangGraph
  • CrewAI

Supporting work

Solvers, experiments and verified exercises — smaller in scope, kept because they show range rather than volume.

  • OptimisationPython · Numba · NumPy +2

    Metaheuristic search over DIMACS benchmark instances

    Benchmarks
    DIMACS suite
    Kernels
    Numba JIT
    Search
    Memetic + hyper-heuristic
    License
    MIT
    Repository
    • Python
    • Numba
    • NumPy
    • NetworkX
  • SystemsC++ · Algorithms · Scheduling

    Competitive scheduling for LLM inference across edge and cloud

    Competition
    ICPC 2026 · Huawei
    Cloud nodes
    1 ≤ K ≤ 8
    Requests
    1 ≤ R ≤ 2000
    Budget
    15s · 256 MB
    Repository
    • C++
    • Algorithms
    • Scheduling
  • IntelligencePython · NumPy · ML Fundamentals

    Core ML algorithms implemented from scratch in NumPy

    Verified
    3 problems
    Stack
    NumPy only
    Source
    Auto-synced
    Repository
    • Python
    • NumPy
    • ML Fundamentals

Repositories that were forked for coursework, or that only mirror a template, are deliberately excluded from this page. Where a project has no published benchmark or screenshot, this site says so rather than inventing one.

04Systems

Three architectures, drawn from the source files.

Diagrams are transcribed from each repository's own service layout, compose file and module tree — not redrawn for effect. Follow any source path to check them.

data flow · left to rightservices/ · infra/docker-compose.yml

Log producers

4 simulated microservices with injected anomaly bursts

Python

Redpanda

Kafka-compatible event bus, dual listeners

Kafka API

Ingestion service

pgx.Batch writes, ON CONFLICT idempotency, embedded migrations

Go

TimescaleDB

Hypertable + log_events_1min continuous aggregate

Postgres

Anomaly engine

Polls every 15s — directional z-score on p95/max, EWMA on error rate

Rust

RAG query service

Natural-language questions grounded in detected anomalies

FastAPI · Ollama

Operator dashboard

Anomaly feed, service health, query terminal

Next.js

Prometheus + Grafana

9-panel dashboard, 15s scrape, committed datasource config

Observability

Events flow left to right; the anomaly engine reaches back into the aggregate to score it, then writes its detections back to the same store.

Languages in one pipeline
4
Go · Rust · Python · TypeScript
Detection cadence
15s
Rust engine over a 1-minute aggregate
Verdict determinism
10/10
AegisTel, deterministic path only

05Build log

Every public repository, forks included.

No curated highlights here — this is the raw public activity, pulled from the GitHub API. Forks are labelled as forks, and repositories without documentation are shown as they are.

Yahia Abdeldjalil Benyahia@yahiaongh
Repos
17
Followers
1
Since
2026
View on GitHub

Repository size by language

Measured in repository kilobytes, original work only. A proxy for where the effort went — not a claim about proficiency.

  • Python32100 KB
  • TypeScript3536 KB
  • Jupyter Notebook1731 KB
  • C++1019 KB
  • HTML303 KB

live from api.github.com03 Nov 2026

api.github.comno contribution-graph proxyhourly revalidation
06Contact

Let's build something that holds up.

Research collaboration, an engineering role, or a system you think I should take apart — all three are interesting. The email address is the fastest route and it reaches me directly.

yahia.infodz@gmail.com

Open to research collaboration & engineering roles

GitHub@yahiaonghLinkedInYahia AbdeldjalilEmailyahia.infodz@gmail.comWhatsApp+213658309707Curriculum vitaePDF · 385 KB
Built with Next.js

Server components for content, client components only where interaction demands it. Every claim on this page is traceable to a repository.

WorkBuild LogContact