Ismail Allouch — AI Systems Engineer

Open to AI engineering internships — 2027
I build retrieval and decisionsystems that know what they don’t know.
AI Systems Engineer
ENSAM Meknès · 2023 — 2027
Meknès, Morocco
Scroll
Grounded answers Grounded answers Grounded answers Grounded answers Grounded answers Grounded answers 
Calibrated abstention Calibrated abstention Calibrated abstention Calibrated abstention Calibrated abstention Calibrated abstention 
Who you would be hiring
I build the parts that have to hold.
Ismail Allouch
Now
Software & Intelligent Systems, ENSAM Meknès — 4th year
Most recent
Engineering intern at Yazaki Morocco, Jul–Sep 2026
Looking for
AI engineering internship, 2027
Works in
Arabic · French · English · Turkish

Retrieval that knows its limits

Hybrid search over a real corpus — lexical and dense, fused — with a router in front and a calibrated refusal behind it.

  • →Recall@1 0.889
  • →15 / 15 out-of-scope refused
  • →0 invented numbers / 82 claims

Agent systems with a cost ceiling

Bounded DAGs instead of open-ended loops, so worst-case latency and spend are known before a request starts.

  • →≤ 4 LLM calls per request
  • →5-stage DAG
  • →8,100+ entries indexed

Shipped, not demoed

Containers, pipelines and TLS on infrastructure I operate. A push to a branch is the whole deploy.

  • →jobgps.ma live
  • →GitLab CI → Oracle Cloud
  • →Docker Compose, Certbot
Selected work
Three systems, all shipped.
01 / 04
01Yazaki MoroccoNDA

Loom

A design guide that answers with the page number — and catches what an engineer would miss.

Catches design-for-manufacture violations at the drawing, not at the plant — where the same fix costs orders of magnitude more. On the reference case it found 14 applicable rules unprompted and flagged two violations a human reviewer had missed.

Rules extracted
110
Recall@1
0.889
F1 gain from gating
+25.5 pts
What this demonstrates
  • Deterministic document extraction (geometry-based, not model-guessed)
  • Hybrid retrieval: BM25 + dense vectors fused with reciprocal rank fusion
  • Calibrated abstention — refusing out-of-scope questions instead of guessing
  • LLM output gating via numeric whitelist
  • Constraint solving with auditable arithmetic
  • Architecture enforced by import-graph tests
Jul — Sep 2026Engineering intern — sole author
Loom architecture: offline ingestion lane and runtime lane
Two lanes. Ingestion runs once and deterministically; runtime never re-reads the PDF.
  • Python
  • FastAPI
  • PyMuPDF
  • BGE-M3
  • FAISS
  • BM25
  • PostgreSQL
  • Ollama
  • Gemini
  • React
  • Docker

DRAG, SWIPE OR USE ← → TO TURN THE PAGE

Toolkit
What I reach for, and how far.

HOVER A BADGE FOR CURRENT LEVEL

92%Python: 92 percent
90%Gemini: 90 percent
90%n8n: 90 percent
88%LangChain: 88 percent
88%Docker: 88 percent
87%Git: 87 percent
85%FastAPI: 85 percent
85%Flask: 85 percent
84%React: 84 percent
84%PostgreSQL: 84 percent
82%GitLab CI: 82 percent
82%Tailwind: 82 percent
80%Node.js: 80 percent
80%Next.js: 80 percent
80%Supabase: 80 percent
79%scikit-learn: 79 percent
77%TypeScript: 77 percent
76%Hugging Face: 76 percent
75%Ollama: 75 percent
OCI74%Oracle Cloud: 74 percent
70%PyTorch: 70 percent
66%Spring Boot: 66 percent
Also built
  • Lead intelligence pipelinesLeanFlow Pro — delivered to client
  • Multi-Agent RAGDocument intelligence
  • 2D-CampusEuromed Fès
  • Agri-GenNo-code IoT
Track record
Where it actually shipped.
Next
Let’s buildsomething that holds.

I’m looking for an AI engineering internship where the systems have to survive contact with real users and real consequences.

GitHubLinkedInCV