I'm an AI graduate based in Dubai. I have an MSc in Artificial Intelligence with Distinction from the University of Hull (2024–2026) and a BSc in Business Management. I speak Arabic (native), English and French.
I design, build with AI coding agents, and evaluate LLM systems for customer-facing work: agents that ask a person before a risky action, multilingual test suites, and assistants that say "I don't know" instead of guessing.
Looking for: AI internships, graduate programmes and entry-level AI roles, remote (US or worldwide) or in Dubai. Available immediately.
- AI agents with guardrails: tool use with LangGraph, identity checks, and human approval for refunds and other risky steps.
- LLM evaluation and red-teaming: Arabic, English and French test sets, baselines, an LLM judge from a different model family, and cost and latency per answer.
- RAG: answers that cite their sources, retrieval measured with hit@k, and an "I don't know" path.
- MCP servers and data agents: read-only database tools and guarded text-to-SQL.
- Voice agents: speech-to-text, then the agent, then text-to-speech, with each stage timed.
- AI governance: risk registers and go/no-go decisions tied to test evidence.
All data is synthetic or public. Every number comes from the saved run files of 8 October 2026 and is explained in the repo's README. Scores marked "judged" come from an LLM judge, not from a person, and I am still reviewing the Arabic and French test items.
| Project | What it does | Headline result | Links |
|---|---|---|---|
| shop-support-agent | Customer-service agent for a fictional skincare shop, in Arabic, English and French. It tracks orders, reads a CRM over MCP and asks a person before any refund. | 97.5% task success on 120 test conversations (rules-only bot: 63.3%), with 0 policy violations | Repo · Live demo |
| mcp-shop-data-agent | A read-only MCP server and an agent that answer business questions in English or Arabic by writing SQL | 48 of 50 questions answered correctly; 39 attack queries led to 0 personal-data leaks and 0 database writes | Repo · Live demo |
| bilingual-rag-assistant | Arabic/English help-centre assistant that cites its sources and says "I don't know" | Embeddings: 44/50 answers judged correct, against 32/50 with keyword search; "I don't know" on 10/10 unanswerable questions | Repo · Live demo |
| clinic-front-desk-assistant | Admin-only clinic assistant in Arabic, English and French. It books appointments and never gives medical advice. | 120/120 emergency messages escalated (a keyword list alone caught 72/120); 0 of 90 attempts got medical advice | Repo · Live demo |
| voice-order-agent | Phone-style order-status agent in Arabic and English. It confirms the order number digit by digit and hands over to a person when needed. | 26/30 test calls succeeded (English 15/15, Arabic 11/15), at about 1.1 s per turn and US$0.006 per call | Repo · Live demo |
| ads-copy-agent | Drafts ad copy in Arabic, English and French and checks it against brand and ad-policy rules. A person approves everything and nothing is posted. | Adding an LLM reviewer to the code rules raised policy-violation recall from 0.67 to 1.00 on 48 held-out ads | Repo · Live demo |
| Project | What it does | Headline result | Links |
|---|---|---|---|
| multilingual-llm-eval | Arabic/English/French evaluation and red-team suite that compares 3 models as a shop's support assistant | 180 quality items and 45 attacks per model: 96.1–100% judged pass rates; 44–45 of 45 attacks blocked | Repo |
| tutor-safety-eval | Go/no-go evaluation of an Arabic/English/French homework tutor: answer leaks under pressure, and escalation when a child is at risk | No setup passed the pre-set gates. The better prompt escalated all 48 harm probes but still leaked answers in 23–35% of pushy dialogues. | Repo · Live demo |
| ai-governance-pack | Risk register, oversight runbook and launch decision for the shop-support agent, with each claim tied to a test | A probe found a gap in the leak filter (3 of 6 variants caught; 6/6 after the fix). Decision: go for a supervised internal pilot, not yet for real customers. | Repo |
| bank-complaints-triage | Routes Arabic, English and French bank messages to 8 teams and flags complaints and card disputes, with an audit trail | LLM + rules routed 92.3% of 300 messages with a 2-point gap between languages (TF-IDF: 47.0%). On English intents, a trained classifier still beat the zero-shot LLM (macro-F1 0.924 vs 0.753). | Repo · Live demo |
| Project | What it does | Headline result | Links |
|---|---|---|---|
| ai-workflow-pack-n8n | Four n8n + AI back-office workflows (enquiries, tickets, a weekly report, invoices), with code checks and human approval | All 8 invoice fields right on 50/50 invoices (regex baseline: 19/50); 0 replies sent without approval | Repo |
| ai-opportunity-assessment | Consulting-style AI assessment: process mining on a public event log, use-case scoring, an ROI model and a pilot with kill criteria | The claim pre-check found 40/40 planted problems (rules alone: 24/40), at US$0.016 per 100 claims | Repo · Live demo |
| ecommerce-analytics-powerbi | 12 business questions on about 100,000 real Olist orders, answered with DuckDB SQL, tests and a decision memo (Power BI dashboard in progress) | 6.8% of orders arrived late; late orders averaged 2.27 stars against 4.29 for on-time orders | Repo |
| Project | What it does | Headline result | Links |
|---|---|---|---|
| NutriAI | .NET MAUI app that recommends recipes from a user's diet, allergies, goals and behaviour, and explains each recommendation (C#, MVVM, SQLite, Spoonacular API). Written by me in 2025–2026; tests and evaluation added in Oct 2026. | 85 passing xUnit tests; a safety evaluation on 30 synthetic users found and documented an allergen-filter gap, with a proposed fix | Repo |
- I start with the spec. Each repo has a
BUILD_SPEC.mdthat sets out the problem, the guardrails, the test set, the baseline to beat and the acceptance tests. - AI coding agents, mainly Claude Code, write most of the code. They generate the code, synthetic data and tests from that spec, and they run the evaluations.
- I review, run and question the results. Each README's "How I used AI agents" section says what the agent generated and what I changed.
- Weak results stay in. They include a missed escalation, an LLM judge marking a correct answer wrong, and Arabic speech-recognition errors. Every number has a model, a date and a denominator.
- NutriAI, my MSc dissertation app (C# / .NET MAUI, SQLite, Spoonacular API), I wrote myself, with limited GitHub Copilot help for boilerplate. It explains each recipe recommendation.
The best way to reach me is on LinkedIn: linkedin.com/in/sara-hebbadj-05309b35a