Sami El Akkad — Applied AI Engineer & Product

Q · 信誉从何而来?

Where does the credibility come from?

Two of China's biggest names. Tsinghua and Baidu — paired, not just one.

institution · 学校

Tsinghua SIGS

清华大学深圳国际研究生院

🥇

AIID Yearly Competition Winner

SSO · NetDragon partner track · 2025

🎓

CSC Scholar · GPA 3.9/4.0

中国政府奖学金 — full scholarship

In Asia

Tsinghua · US News 2024–25

MSc Artificial Intelligence · AI+X Innovation Design track · Shenzhen · 2025–Present

ernie mentor program · 项目

Baidu ERNIE

百度文心导师 · Wenxin Mentor

🥉

3rd Prize · ERNIE Hackathon 2025

RedFox · Hardware Track · top 3 of 23 finalists · 170+ subs · ~5,900 devs

💼

Paid contract role

Data construction · eval rubric · prompts

📜

Two official Baidu certificates

ERNIE Mentor Program · 2025

ERNIE / 文心一言 · Shenzhen · Oct–Dec 2025

网龙

third signal · 雇主管道

NetDragon Websoft (网龙网络) — Fuzhou employer pipeline

Fortune-China AI + Education leader · AIID partner. SSO was built for their stack. AIID Visiting Study (Fuzhou, Dec 2025) opens the introduction.

🥇

scholarship · 奖学金

NPU President's Full Prize

100% tuition + ¥15,000 stipend

⚡

selective track · 精英计划

AI+X Elite Program

Tsinghua SIGS

🎓

language · 中文

HSK 4 · 90/100

NPU transcript verified

📐

methodology · 方法论

5-Dim Eval Rubric

DOPEDCLUB → Baidu → jak.ma

Q · 数字一览

By the numbers

A quick snapshot — every project below shows its real output (detections, dashboards, proofs that run).

green CI builds ✓

public repos

AI internships

languages spoken

Q · 为什么这样建?

Why I build the way I do.

"Most AI products treat low-resource languages and emerging markets as English-first with a translation layer on top. I think that's the wrong abstraction, and the systems I build try to prove it."

My way of thinking about AI products came out of my undergrad thesis at NPU (ADCMOplus · with Prof. Lin Li) on constrained dynamic multi-objective optimization. AI products almost never come down to a single metric — you're trading latency, cost, accuracy, and fairness against each other, and the honest answer is a Pareto front, not a winner. The six-check verifier in jak.ma is the same reasoning applied at runtime.

Practically, this means I work in Arabic, Darija, French, English, and functional Mandarin — I can ship products into MENA, work inside a Chinese tech stack, and connect the two without the usual translation cost.

Q · 谁报道过我?

Who has officially featured me?

Tsinghua SIGS itself — part of the official #MyChinaStory + #TsinghuaRen series. Published on @tsinghua.sigsadm · Feb 2026. official

View on Instagram →

"Through his studies and work on AI projects, including his involvement in the Baidu ERNIE ecosystem, he gained firsthand insight into what he describes as 'a fast-paced tech ecosystem focused on execution, rapid iteration, and real-world impact.'"

— Tsinghua SIGS (@tsinghua.sigsadm) · #MyChinaStory · Feb 2026

Tsinghua SIGS Story — defining moment: AIID first prize, Baidu ERNIE National Competition, Tencent, NetDragon

defining moment · AIID 1st · Baidu ERNIE · Tencent · NetDragon ↗

Tsinghua SIGS Story — what working with Baidu taught me about innovation

what Baidu taught me about innovation ↗

My Tsinghua SIGS Story — Sami El Akkad, AI + Innovation Design, CSC Scholar

My Tsinghua SIGS Story · cover ↗

Tsinghua SIGS Story — on Shenzhen and tech integrated into daily life

on Shenzhen · tech in daily life ↗

Tsinghua SIGS Story — on Tsinghua's international community

on the international community ↗

Tsinghua SIGS Story — advice: engage globally to accelerate growth

advice · go global to grow ↗

Tsinghua SIGS Story — how internships and hands-on experience won the CSC scholarship

how I earned the CSC scholarship ↗

more than a degree · closing ↗

also covered by · 也被报道

@Tsinghua_Uni · verified main account.

Not just SIGS — the central Tsinghua University verified Twitter picked up the story.

Tsinghua University verified tweet about Sami — OpenFIESTA + Shenzhen tech ecosystem

at Tsinghua OpenFIESTA · "fast-paced ecosystem · execution · iteration · measurable impact" ↗

Tsinghua University verified tweet about Sami — Three Cultures One Mission, Chinese New Year integration

Three Cultures, One Mission · Chinese New Year · cultural integration ↗

Q · 我的专业是什么?

What's the AI+X program I'm in?

Tsinghua SIGS AI+X · Innovation Design. Four core domains, applied to ten verticals, with the industry partners who actually hire the graduates.

core domains · 核心方向

The four pillars of AI+X

Smart Network
Automation

Core
Algorithm

Electronic
Communication

Computer
Technology

empowering applications · AI+X 赋能应用

Applied to ten verticals

🏭 Manufacturing

💰 Finance

🚦 ITS / Transport

🩺 Healthcare

🏙️ Smart City

📊 Business Intelligence

🌊 Smart Ocean

🌐 Smart Network

🥽 Multi-verse

📡 Software & Hardware

industry partners · 行业伙伴

Who hires AI+X graduates

UBTECH 优必选 · Huawei 华为 · ByteDance 字节跳动 · Tencent 腾讯 · Ping An 平安科技 · SenseTime 商汤 · Meituan 美团 · Alibaba Cloud 阿里云 · Sangfor 深信服 · Parametrix.ai

AI+X Innovation Design is a selective Tsinghua SIGS track combining core AI domains with industry verticals — partnered directly with the Chinese AI ecosystem.

Q · 我究竟做过什么?

What have I actually built?

From shipped products to from-scratch models. One flagship marketplace, then everything else in four bands — products, models, evaluation, research. 16 pieces of work · 12 public repos · 6 with passing CI ✓. All rights reserved (contact to reuse).

github.com/Samielakkad →

live in production founder

jak.ma · Moroccan home-services marketplace

Zero-commission, anti-shnaq, Darija-first chat marketplace for skilled-trade workers (plumbers, electricians, tilers). 1,862 workers · 15 cities · 3,400 daily queries · 1,100 DAU. Two-pass architecture: Pass 1 classifies intent + city, Pass 2 generates a constrained response that must clear a six-check verifier before reaching the user. p50 ≈ 1.2s · ≈$0.0008 per query · <1% verifier rejection in production. The price-band check is built from a 200-worker field survey — there's no published price reference for Moroccan trades, and an LLM-as-judge would just rubber-stamp the model's biased prior. Beats sabab.ma (French-first incumbent) on worker count (4× more).

JS · Node · Mongo grok-3-mini verifier-gated

01 · Live products & apps

🥇 AIID Winner ▶

Sovereign School OS

TypeScript · ✓ CI · NetDragon 网龙

AI-native K-12 platform — auth · real-time chat · AI tutoring on Wenxin (ERNIE) · parent dashboard. 50+ React/TS components.

▶ Watch demo Code ↗

🎨 Best Design

Sinothrive AI

2025 · Tsinghua · 6-member team

AI assistant matching international students to universities in China. Won Best Design Award, Dec 2025.

🥉 Baidu · 3rd Prize

RedFox · ERNIE Hackathon

2025 · Jetson AGX Xavier

AI hardware companion on Wenxin (ERNIE). Top 3 / 23 finalists · 170+ subs · ~5,900 devs. SFT won 87.5% / 72.5% (Interactivity / Safety), blind A/B.

⚖️

legal-tech

✓ CI

mirath-morocco

TypeScript · ✓ CI

Moudawana inheritance + land-registry solver — Moroccan family law, productized.

02 · Models, vision & from-scratch ML

✓ CI

YOLOv8 from scratch

Python · @ Neusoft 东软 · 2024

Autonomous-driving road-object detector built from zero in PyTorch — CSPDarknet · PANet · DFL head. Tests + Dockerfile + CI.

✓ CI

ADCMOplus

Python · @ NWPU 西北工大 · 2025

Online-learning evolutionary algorithm for dynamic constrained optimization. Runnable Theorem-1 regret demo. Rank 1.25 vs 5 SOTA.

🗣️

HF model

darija-classifier

Python · Hugging Face

Production Moroccan-Arabic intent + city classifier powering jak.ma's Pass-1.

🤗

live demo

🤗 live

Darija Chat — HF Space

2026 · Hugging Face Spaces

Interactive Darija chat demo — production retrieval architecture, browser-runnable. Public, deployed, reproducible.

03 · Evaluation & PM methodology

methodology

pm-frameworks-darija

Markdown · 5 files

Pricing · eval · calibration · verifier · multimodal. Rater calibration to Krippendorff α ≥ 0.7. The layer behind every project here.

5-dimension evaluation radar — factuality, reasoning, fluency, calibration, safety

jak-ma-eval-suite

Python

5-dim evaluation rubric for two-pass grounded retrieval — the verifier that gates jak.ma in production.

📏

Baidu 百度

ernie-evaluation-notes

Markdown · Baidu provenance

5-dim eval (factuality · reasoning · fluency · calibration · safety) · adversarial pairs · cross-language calibration. Feeds release-gating.

jak.ma production metrics — 1,862 workers, 3,400 daily queries, p50 1.2s, $0.0008/query, <1% verifier rejection

jak-ma-case-study

Markdown · ~6,500 words

Production case study — 5 decisions · 5 things that broke · the numbers a year in.

04 · Research, docs & resources

🧬

thesis · NPU

thesis-adcmoplus

MATLAB · inline write-up

74-page NPU undergrad thesis · Friedman rank 1.25. Full write-up on this page →

aiid-team-docs

TypeScript · ✓ CI

Interactive docs site — Best Documentation Award, Tsinghua AIID Hackathon 2025.

🗂️

resources

darija-nlp-resources

Markdown

Curated Moroccan-Arabic NLP resources — datasets, papers, tooling.

Q · 我的研究是什么?

My undergrad thesis — ADCMOplus.

74 pages. Rank #1 of 6 algorithms tested. A novel evolutionary algorithm that classifies the type of environmental change (objective / constraint / combined) before adapting — built for problems where targets and constraints both move over time. Maps onto LLM serving, RLHF, agent orchestration.

1.25

Friedman rank · best of 5 SOTA

12 / 12

benchmarks won on diversity (Wilcoxon p<0.05)

10 / 12

benchmarks won on DeltaP

pages · MATLAB R2024B

May 2025

defended · NPU · Li Lin

Read the full thesis write-up (PM-framed)

Contributions · architecture · benchmark results · applications to LLM serving, RLHF, agent orchestration · ~6,000 words

Expand ↓

Underlying MATLAB code + the 74-page PDF stay private — email sam25@mails.tsinghua.edu.cn for access.

Q · 我在哪里工作过?

Where have I worked?

Four AI roles in 18 months. The thread: one 5-dim evaluation methodology, invented at DOPEDCLUB, formalised at Baidu ERNIE.

Oct 2025 – Dec 2025 · Shenzhen

paid contract

Baidu — ERNIE Mentor

百度 · 文心导师 · Wenxin Mentor

Paid expert on Baidu's flagship LLM (文心一言 / ERNIE)
Stress-tested prompts · multi-dim scoring rubric · production-readiness signals
Drove release decisions on candidate checkpoints
Built RedFox AI hardware companion on Wenxin + Jetson AGX Xavier
🥉 3rd Prize · top 3 / 23 finalists · 170+ subs · ~5,900 devs · Hardware Track recap feature
Two official Baidu certificates

May 2025 – Jul 2025 · Hebei

industrial ML

Hebei Chinaust Plastics — IT Engineer (AI & Robotics)

ML predictive maintenance across 3 production lines
+12% accuracy · −18% downtime
Data pipeline on 120K+ sensor records
Digital twins cut product test cycles 2 weeks → 4 days
Authored Morocco-expansion feasibility study (~30% logistics savings)

Sep 2024 – Feb 2025 · Casablanca (Hybrid)

applied AI-PM

DOPEDCLUB — Applied AI Product Manager

Creative studio using GenAI for client visual deliverables
Designed the 5-criteria evaluation framework — origin of the signature rubric
Same rubric was later reused at Baidu ERNIE and as the jak.ma verifier
12-item checklist/SOP across 20+ recurring client briefs
Longest-duration formal AI-PM role on the CV

Jun 2024 – Sep 2024 · Zhengzhou-Kaifeng

CV / autonomous-driving

Neusoft — AI Intern

Object detection for autonomous-driving R&D · ~98% precision
Segmentation pipelines: −45% preprocessing time · −30% training time
Healthcare + e-commerce segmentation deployments
Identity-verification eval supporting internal Alipay / WeChat Pay integration tests

Q · 我是怎么走到这里的?

How did I get here?

NPU undergrad → Tsinghua master's. CSC government scholar. Five languages.

2025 — Present · Shenzhen

current

Tsinghua SIGS — MSc Artificial Intelligence (AI+X Innovation Design track)

清华大学深圳国际研究生院 · CSC Government Scholar

2021 — 2025 · Xi'an

Northwestern Polytechnical University — B.Eng. CS

西北工业大学 · Thesis defended May 2025 · 74-page ADCMOplus

NPU President's Scholarship — Full Prize: 100% tuition waiver + ¥15,000 stipend

languages

· Arabic · native

· Darija · native

· English · fluent

· French · fluent

· Mandarin 中文 · HSK 4 (90/100, NPU transcript)

Q · 我怎么做产品?

How do I think about AI products?

The honest answer is a Pareto front, not a winner. Classify the failure mode first, then pick the right tool.

AI products almost never come down to a single metric — you're trading latency, cost, accuracy, and fairness against each other. That's the through-line from my undergrad thesis (ADCMOplus · constrained dynamic multi-objective optimization) to the runtime architecture of jak.ma.

My signature methodology — 5-dim evaluation rubric + verifier-gated retrieval — started at DOPEDCLUB (Casablanca), got formalised at Baidu ERNIE (Shenzhen), and now runs in production at jak.ma (Morocco) as a six-check verifier.

pricing taxonomy

evaluation rubric

calibration protocol

verifier philosophy

multimodal decision

Q · 我最近在读什么?

Currently exploring.

The frontier work shaping how I think. What I'm reading, building toward, and arguing with.

alignment · safety

Constitutional AI · RLHF

Anthropic's CAI work and the broader RLHF stack — the multi-reward shaping problem in ADCMOplus's change-classification framing.

retrieval · agents

Tool-use & verifier-gated agents

OpenAI / Anthropic tool-use, Manus, multi-agent orchestration — directly tied to jak.ma's two-pass architecture.

evaluation

LLM-as-Judge · HELM · LMSys

How major labs validate models. Extending the 5-dim rubric from Baidu ERNIE work toward benchmark-grade reproducibility.

low-resource NLP

Darija + Arabic dialect LLMs

Atlas Chat, Jais, Falcon, AceGPT — Maghreb dialect coverage and the gap between MSA and real spoken Arabic.

fine-tuning

LoRA + small-model fine-tuning

Qwen2.5 / Llama 3.1 / ERNIE-Tiny LoRA recipes. Roadmap: 3× cost reduction on jak.ma's Pass 2 generation.

china frontier

ERNIE · DeepSeek · Qwen · Moonshot

The Chinese frontier-model landscape — release cadence, eval gaps, where each beats GPT/Claude on price-per-eval.

Q · 我在找什么?

What am I looking for?

Applied AI / ML engineering & product roles starting 2026 — open from a 3-month internship through a full-time start.

I care about the team and the problem more than the logo or the location. If you're building something hard in applied AI and think I'd fit, I want to hear from you — wherever you are.

Some ecosystems I already know how to operate in — examples, not a checklist:

Chinese tech going international (出海) — ByteDance / TikTok, Tencent, Alibaba, Xiaomi, SHEIN, Lenovo, Huawei & peers.
Global labs & firms with a China presence — e.g. Microsoft Research Asia, NVIDIA, Apple.
Southeast Asia & Singapore — Sea AI Labs, AI Singapore, GovTech, Grab, Lazada & others.
Gulf & MENA AI — G42, Inception, MBZUAI, TII Falcon, HUMAIN & the wider region.
…or anywhere else — remote, Europe, North America, back home in Morocco — a great team is a great team.

Drawn to (but not limited to): AI + Education · AI hardware · multi-agent retrieval · low-resource-market AI (Arabic / Darija / Mandarin / French) · evaluation & calibration · marketplace / matching. Based in Shenzhen · open to relocation and remote.

sam25@mails.tsinghua.edu.cn linkedin.com/in/samielakkad github.com/Samielakkad 🤗 jakma-darija-chat (live)