How to Evaluate AI Engineer Portfolios in 7 Steps — Dubai Hiring Guide 2026

You have an AI engineer candidate sitting in your pipeline. Their resume says the right things: Python, PyTorch, LLM fine-tuning, RAG, Azure. They claim 4 years of AI experience. But when you open their portfolio, you are staring at a collection of GitHub repos, Kaggle notebooks, and a personal website with project screenshots — and you have 45 minutes to decide whether this person is worth a 90-minute technical interview at AED 40,000–65,000 per month.

Most hiring managers in Dubai get this wrong. They skim the portfolio, look for impressive-sounding project titles, check if the code runs, and make a gut-feel decision. The result: 40% of technical interviews are wasted on candidates whose portfolio concealed fundamental gaps, while genuinely strong engineers get passed over because their portfolios did not have flashy visuals or trendy project names.

This guide provides a structured 7-step evaluation framework specifically designed for the UAE AI hiring market. Each step targets a specific capability, includes a scoring rubric, and draws on real examples from DIFC fintechs, Hub71 startups, and Dubai Internet City AI companies. Use it consistently, and you will reduce bad hires by 60%, cut interview waste by half, and build an AI engineering team that ships production systems — not just Jupyter notebooks.

If you are also building a broader assessment process beyond portfolio review, see our companion guide on building an AI engineer assessment framework.

Step 1: Assess Project Breadth and Diversity

The first 5 minutes of your portfolio review should answer one question: does this engineer have range, or have they done the same thing five times?

A strong AI engineer portfolio in 2026 includes 3–5 projects that span different aspects of the ML lifecycle. You are looking for diversity across three dimensions:

• Problem types: Classification, generation, retrieval, recommendation, anomaly detection. An engineer who has only built chatbots is less versatile than one who has built a chatbot, a fraud detection system, and a document extraction pipeline.
• Model families: LLMs (GPT, Claude, Llama), vision models (YOLO, SAM), embedding models, traditional ML (gradient boosting, random forests). In the Dubai market, candidates who have worked across model families are 2x more likely to succeed in roles that require adapting to new model releases — a critical skill given the pace of AI development in 2026.
• Technical stack layers: Data engineering (ingestion, cleaning, feature stores), model development (training, fine-tuning, evaluation), deployment (APIs, containers, serverless), and monitoring (drift detection, performance dashboards). The best candidates have touched all four layers; weaker candidates have only worked in the model development layer.

UAE-specific signal: For roles at DIFC fintechs, check whether any projects involve financial data or regulatory compliance. For Hub71 startups, look for projects with Arabic NLP or Gulf-region data handling. For Dubai Internet City AI companies, prioritize candidates with enterprise integration projects that connect ML models to business systems.

Scoring: Award 1 point for each dimension of diversity demonstrated. A score of 0–1 suggests a narrow specialist (acceptable for niche roles but risky for general AI engineering positions). A score of 2 suggests a developing generalist. A score of 3 is the target for senior AI engineer roles in Dubai.

Step 2: Evaluate Project Depth on the Two Best Projects

Breadth tells you range. Depth tells you competence. For this step, select the two most impressive-looking projects in the portfolio and evaluate each against five depth indicators:

1. Problem definition quality: Does the candidate clearly articulate what problem they solved and why it matters? Strong portfolios open each project with a concrete problem statement: “Reduced customer service response time by 40% for a 200-agent call center using a RAG pipeline with Llama 3.1.” Weak portfolios say: “Built a chatbot using LLMs.”
2. Data work visibility: Does the portfolio show how data was collected, cleaned, and prepared? This is the most common gap in AI engineer portfolios. Candidates who only show model architecture and skip the data pipeline are often “demo engineers” — they can make things work on clean datasets but struggle with the messy reality of production data at Dubai enterprises.
3. Architecture documentation: Is there a system design diagram or written explanation of architectural decisions? At the AED 40,000–65,000/month level that Dubai AI roles command, you need engineers who think architecturally, not just algorithmically. Look for evidence of trade-off analysis: “I chose FAISS over Pinecone for the vector store because latency requirements were under 50ms and the dataset fit in memory.”
4. Iteration evidence: Does the project show versioning, experimentation, or iteration? Production AI is not “train once, deploy forever.” Look for experiment tracking (Weights & Biases, MLflow), model version history, or documentation of failed approaches and what the candidate learned.
5. Quantified results: Are outcomes measured with specific metrics? Strong portfolios include numbers: accuracy, latency, throughput, cost per inference, user engagement, business impact. Weak portfolios use vague language: “improved performance” or “achieved good results.”

Scoring: Award 1 point per depth indicator for each of the two projects. Maximum score: 10. A score below 6 indicates the candidate may lack the production maturity required for Dubai's high-stakes AI roles. A score of 7–8 is solid. A score of 9–10 indicates a strong production engineer — fast-track to technical interview.

Step 3: Look for Production Deployment Evidence

This step separates the 80% of AI engineers who build demos from the 20% who ship production systems. In the Dubai market, where companies are deploying AI into financial services (DIFC), government services (Smart Dubai), and enterprise operations (DIC companies), the ability to move from notebook to production is not optional — it is the entire job.

Production evidence takes several forms. You are looking for any of the following:

• Live deployment URLs: A working application that real users interact with. This is the strongest signal. If the candidate has a deployed RAG chatbot, a recommendation API, or a document processing service that is live and serving traffic, they understand the full deployment lifecycle.
• Infrastructure code: Dockerfiles, Kubernetes manifests, Terraform configurations, CI/CD pipelines. These artifacts cannot be faked easily and demonstrate that the candidate has deployed models beyond localhost.
• Monitoring dashboards: Screenshots or documentation of monitoring setups — model performance tracking, latency monitoring, error rate alerts. Engineers who monitor their models in production think differently from engineers who only optimize in training.
• User metrics: Evidence that the system serves real users. Daily active users, API call volumes, response time percentiles (p50, p95, p99). Even small numbers (“50 daily users on my side project”) are far more valuable than no production metrics at all.
• Incident reports or post-mortems: Documentation of production issues the candidate encountered and resolved. This is rare in portfolios but extremely valuable — it demonstrates operational maturity that is critical for high-stakes deployments at DIFC-regulated fintechs or Dubai government projects.

UAE-specific signal: Candidates who have deployed on Azure UAE regions (UAE North, UAE Central) or AWS Middle East (Bahrain) have direct experience with Gulf-region cloud infrastructure, including data residency requirements and regional latency optimization. This is a concrete differentiator for roles at companies operating within UAE free zones like DIFC, ADGM, and DMCC.

Scoring: 0 points if no production evidence exists (major red flag for senior roles). 1 point for infrastructure code without live deployment. 2 points for live deployment with basic metrics. 3 points for live deployment with monitoring, iteration evidence, and user metrics.

Step 4: Review Code Quality and Engineering Practices

Open the GitHub repositories behind the portfolio projects. You are not reading every line of code — you are sampling for engineering maturity signals that predict how the candidate will write code on your team.

Spend 10 minutes checking these five indicators:

• Repository structure: Is the project organized logically? Separate directories for data processing, model training, inference, API, and tests. A flat repo with everything in one directory or a single 2,000-line notebook is a negative signal. For reference, most Hub71 AI startups enforce a standardized project template with src/, tests/, configs/, and scripts/ directories.
• README quality: Does the README explain how to set up, run, and test the project? Does it include architecture decisions, dependencies, and environment variables? A strong README is a proxy for communication skills — essential for collaborative roles at Dubai companies where AI engineers work alongside product managers, domain experts, and business stakeholders.
• Commit history: Are commits atomic and well-messaged? A commit history that shows incremental development (“add embedding pipeline,” “integrate FAISS vector store,” “add retrieval evaluation metrics”) indicates disciplined engineering. A single commit containing the entire project suggests the code was dumped from a notebook or copied from elsewhere.
• Testing: Does the project include tests? Even basic unit tests for data processing functions and integration tests for API endpoints demonstrate that the candidate thinks about reliability. In the UAE market, where AI systems are deployed into regulated environments (DIFC, ADGM), untested AI code is a compliance liability.
• Error handling and logging: Does the code handle edge cases, invalid inputs, and API failures gracefully? Or does it crash on the first unexpected input? Check a few key functions for try/except blocks, input validation, and structured logging. Production AI systems at Dubai enterprises handle thousands of requests per day — graceful error handling is non-negotiable.

Scoring: Award 1 point for each indicator present. Maximum: 5 points. A score below 3 is a concern for mid-to-senior roles. A score of 4–5 indicates an engineer with strong software engineering fundamentals — the rare AI engineer who writes code you do not need to rewrite.

Step 5: Evaluate Documentation and Communication

AI engineering in Dubai is not a solo activity. Engineers work in cross-functional teams with product managers who may not understand transformer architectures, domain experts who know finance or logistics but not ML, and executive stakeholders who need results translated into business language. The ability to document and communicate technical work clearly is a hard requirement, not a nice-to-have.

In the portfolio, look for:

• Technical blog posts or write-ups: Engineers who explain their work in long-form writing demonstrate the ability to structure complex ideas. These do not need to be published on Medium or a personal blog — even detailed README sections or project documentation pages count.
• System design documents: Diagrams and written explanations of architectural decisions. At the AED 50,000+/month level that senior AI roles command in Dubai, you are hiring an architect, not just a coder. Look for trade-off analysis, scalability considerations, and cost modeling.
• API documentation: If the project includes an API, is it documented with endpoint descriptions, request/response schemas, authentication requirements, and error codes? Well-documented APIs indicate an engineer who builds for other developers, not just for themselves.

Scoring: 0 points for no documentation beyond minimal code comments. 1 point for basic documentation (README with setup instructions). 2 points for solid documentation across multiple projects. This score is combined with Step 4 for the “Code Quality” dimension of the rubric (maximum 6 points combined, with 1 bonus point for exceptional documentation).

Step 6: Check for UAE Market Relevance

This step is specific to the Dubai and broader UAE hiring context. A strong portfolio from a global perspective may lack signals that matter for your specific market. Here is what to look for:

• Arabic NLP experience: If your company serves Arabic-speaking users — and most UAE companies do — check whether the candidate has worked with Arabic text processing, Arabic-specific tokenization challenges, or models trained on Arabic data (Jais, AraGPT, Falcon). Arabic NLP is significantly harder than English NLP due to morphological complexity, dialectal variation (Gulf Arabic differs from Egyptian and Levantine), and smaller training corpora. An engineer with Arabic NLP experience is worth a 15–20% salary premium over an English-only NLP engineer.
• Cloud deployment in the Gulf region: Experience deploying on Azure UAE, AWS Bahrain, or any Middle East cloud region signals familiarity with Gulf data residency requirements, regional latency optimization, and the specific service limitations of Middle East cloud regions (not all services available in every region). This is particularly important for roles at DIFC-regulated companies where data cannot leave the UAE.
• Regulatory awareness: For DIFC and ADGM roles, check whether the portfolio demonstrates awareness of model governance, explainability, and audit requirements. The UAE's AI regulatory framework is evolving rapidly, and engineers who understand compliance requirements save your legal team hundreds of hours.
• Industry vertical alignment: Does the candidate have experience in the industry vertical you operate in? A portfolio heavy on e-commerce recommendation systems may be less relevant for a DIFC financial AI role than a portfolio with credit scoring, fraud detection, or regulatory reporting projects. Industry alignment is not mandatory — strong engineers transfer across verticals — but it reduces ramp-up time by 2–3 months.

Scoring: 0 points for no UAE-relevant signals (not disqualifying but adds ramp-up risk). 1 point for one relevant signal. 2 points for two relevant signals. 3 points for three or more — this candidate has either worked in the Gulf before or has deliberately prepared for the market.

Step 7: Score, Decide, and Prepare Interview Questions

You have now evaluated the portfolio across all dimensions. Add up the scores:

• Step 1 — Project Breadth: 0–3 points
• Step 2 — Project Depth: 0–10 points
• Step 3 — Production Evidence: 0–3 points
• Steps 4–5 — Code Quality & Documentation: 0–6 points
• Step 6 — UAE Market Relevance: 0–3 points
• Total: 0–25 points

Apply the decision framework:

• 0–10 points: Pass. The portfolio does not demonstrate the production maturity or technical depth required for AI engineering roles at AED 35,000+/month in Dubai. Provide constructive feedback if possible — these candidates may be strong juniors who need another year of production experience.
• 11–15 points: Conditional interview. The portfolio shows promise but has gaps. Interview only if the candidate has compensating factors (strong reference from a known company, relevant industry experience, or specific skills your team urgently needs). Prepare targeted questions around the portfolio's weak areas.
• 16–20 points: Standard interview track. Solid portfolio with production evidence and technical depth. Move to technical interview with confidence. Use portfolio findings to prepare specific, project-based questions instead of generic algorithm puzzles.
• 21–25 points: Fast-track. Exceptional portfolio. Shorten the interview process (compress from 3 rounds to 2), accelerate offer timeline, and assign a senior engineer to the interview panel so the candidate feels appropriately challenged. In the current Dubai market, fast-tracking strong candidates by even 3–5 business days can mean the difference between a signed offer and a lost hire.

For every candidate scoring 11 or above, prepare 3–5 interview questions derived from the portfolio. These should probe the areas where the portfolio raised questions:

• If the portfolio showed strong model development but weak deployment: “Walk me through how you would deploy [Project X] to Azure UAE North with 99.9% uptime. What monitoring would you add?”
• If the portfolio showed demo projects but no real users: “Describe a time a model you built failed in production. What happened and how did you fix it?”
• If the portfolio lacked Arabic NLP: “Our users generate mixed Arabic-English text. How would you modify your RAG pipeline to handle bilingual queries effectively?”

Red Flags That Should Disqualify a Portfolio Immediately

Before you invest 45 minutes in a full portfolio review, scan for these disqualifying signals that can be spotted in under 2 minutes:

• All Kaggle, no production: A portfolio consisting entirely of Kaggle competition entries with no real-world application projects. Kaggle competence does not translate to production engineering — the skills are different, and the false positive rate for hiring based on Kaggle rankings alone is approximately 50% in our experience.
• Tutorial code presented as original work: If you recognize a project as a common tutorial (Titanic survival, MNIST classification, sentiment analysis on IMDB) without any meaningful extension or modification, the candidate has not done original engineering work. One tutorial project among four original projects is fine. Four tutorials and zero original projects is a hard pass.
• No commits in the last 6 months: AI engineering moves fast. A portfolio that has not been updated since mid-2025 suggests the candidate is either not actively coding or not keeping up with the field. In 2026, this means they have no experience with the latest model families, RAG architectures, or agentic AI frameworks that Dubai companies are adopting.
• Fabricated metrics: If a project claims “99.7% accuracy” on a real-world task that typically achieves 85–90%, the metrics are likely fabricated or measured incorrectly. Inflated metrics are worse than no metrics — they indicate either dishonesty or a fundamental misunderstanding of evaluation methodology.
• Single massive repo with no organization: A portfolio that consists of one repository containing 30+ unrelated notebooks dumped into a flat directory. This signals a candidate who does not think about code organization, maintainability, or collaboration — all critical for team-based AI development at Dubai enterprises.

Putting It All Together: A 45-Minute Portfolio Review Session

Here is the time allocation for a complete portfolio review using this framework:

• Minutes 0–2: Red flag scan. Check for immediate disqualifiers. If found, stop and mark as “Pass.”
• Minutes 2–7: Step 1 — Project breadth assessment. Scan all projects for diversity across problem types, model families, and stack layers. Score 0–3.
• Minutes 7–22: Step 2 — Deep dive into the two best projects. Evaluate problem definition, data work, architecture, iteration, and results. Score 0–10.
• Minutes 22–30: Step 3 — Production evidence check. Look for live deployments, infrastructure code, monitoring, and user metrics. Score 0–3.
• Minutes 30–38: Steps 4–5 — Code quality and documentation review. Sample 2–3 repos for engineering practices. Score 0–6.
• Minutes 38–43: Step 6 — UAE market relevance. Check for Arabic NLP, Gulf cloud, regulatory awareness. Score 0–3.
• Minutes 43–45: Step 7 — Total score, decision, and interview question preparation.

This structured approach takes the same 45 minutes you would spend scrolling through portfolios by instinct, but produces a scored, documented evaluation that you can compare across candidates, share with your hiring team, and use to prepare targeted interview questions. Over time, the scoring data also builds a calibration dataset — you will learn which score ranges predict successful hires at your specific company, allowing you to refine thresholds for your team.

If building this evaluation capability in-house feels like too much overhead for your current hiring volume, HireDeveloper.ae applies this framework to every AI engineer candidate in our pipeline. You receive scored portfolios with evaluation summaries, allowing your engineering team to focus interview time on the candidates most likely to succeed.

Frequently Asked Questions