Building AI-Powered Recommendation Systems

Last week, I described four design patterns for AI agentic workflows that I believe will drive significant progress: Reflection, Tool use, Planning and Multi-agent collaboration. Instead of having an LLM generate its final output directly, an agentic workflow prompts the LLM multiple times, giving it opportunities to build step by step to higher-quality output. Here, I'd like to discuss Reflection. It's relatively quick to implement, and I've seen it lead to surprising performance gains. You may have had the experience of prompting ChatGPT/Claude/Gemini, receiving unsatisfactory output, delivering critical feedback to help the LLM improve its response, and then getting a better response. What if you automate the step of delivering critical feedback, so the model automatically criticizes its own output and improves its response? This is the crux of Reflection. Take the task of asking an LLM to write code. We can prompt it to generate the desired code directly to carry out some task X. Then, we can prompt it to reflect on its own output, perhaps as follows: Here’s code intended for task X: [previously generated code] Check the code carefully for correctness, style, and efficiency, and give constructive criticism for how to improve it. Sometimes this causes the LLM to spot problems and come up with constructive suggestions. Next, we can prompt the LLM with context including (i) the previously generated code and (ii) the constructive feedback, and ask it to use the feedback to rewrite the code. This can lead to a better response. Repeating the criticism/rewrite process might yield further improvements. This self-reflection process allows the LLM to spot gaps and improve its output on a variety of tasks including producing code, writing text, and answering questions. And we can go beyond self-reflection by giving the LLM tools that help evaluate its output; for example, running its code through a few unit tests to check whether it generates correct results on test cases or searching the web to double-check text output. Then it can reflect on any errors it found and come up with ideas for improvement. Further, we can implement Reflection using a multi-agent framework. I've found it convenient to create two agents, one prompted to generate good outputs and the other prompted to give constructive criticism of the first agent's output. The resulting discussion between the two agents leads to improved responses. Reflection is a relatively basic type of agentic workflow, but I've been delighted by how much it improved my applications’ results. If you’re interested in learning more about reflection, I recommend: - Self-Refine: Iterative Refinement with Self-Feedback, by Madaan et al. (2023) - Reflexion: Language Agents with Verbal Reinforcement Learning, by Shinn et al. (2023) - CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing, by Gou et al. (2024) [Original text: https://lnkd.in/g4bTuWtU ]

BREAKING! The FDA just released this draft guidance, titled Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations, that aims to provide industry and FDA staff with a Total Product Life Cycle (TPLC) approach for developing, validating, and maintaining AI-enabled medical devices. The guidance is important even in its draft stage in providing more detailed, AI-specific instructions on what regulators expect in marketing submissions; and how developers can control AI bias. What’s new in it? 1) It requests clear explanations of how and why AI is used within the device. 2) It requires sponsors to provide adequate instructions, warnings, and limitations so that users understand the model’s outputs and scope (e.g., whether further tests or clinical judgment are needed). 3) Encourages sponsors to follow standard risk-management procedures; and stresses that misunderstanding or incorrect interpretation of the AI’s output is a major risk factor. 4) Recommends analyzing performance across subgroups to detect potential AI bias (e.g., different performance in underrepresented demographics). 5) Recommends robust testing (e.g., sensitivity, specificity, AUC, PPV/NPV) on datasets that match the intended clinical conditions. 6) Recognizes that AI performance may drift (e.g., as clinical practice changes), therefore sponsors are advised to maintain ongoing monitoring, identify performance deterioration, and enact timely mitigations. 7) Discusses AI-specific security threats (e.g., data poisoning, model inversion/stealing, adversarial inputs) and encourages sponsors to adopt threat modeling and testing (fuzz testing, penetration testing). 8) And proposed for public-facing FDA summaries (e.g., 510(k) Summaries, De Novo decision summaries) to foster user trust and better understanding of the model’s capabilities and limits.

🔮 Design Patterns For AI Interfaces (https://lnkd.in/dyyMKuU9), a practical overview with emerging AI UI patterns, layout considerations and real-life examples — along with interaction patterns and limitations. Neatly put together by Sharang Sharma. One of the major shifts is the move away from traditional “chat-alike” AI interfaces. As Luke Wroblewski wrote, when agents can use multiple tools, call other agents and run in the background, users orchestrate AI work — there’s a lot less chatting back and forth. In fact, chatbot widgets are rarely an experience paradigm that people truly enjoy and can fall in love with. Mostly because the burden of articulating intent efficiently lies on the user. It can be done (and we’ve learned to do that), but it takes an incredible amount of time and articulation to give AI enough meaningful context for it to produce meaningful insights. As it turned out, AI is much better at generating prompt based on user’s context to then feed it into itself. So we see more task-oriented UIs, semantic spreadsheets and infinite canvases — with AI proactively asking questions with predefined options, or where AI suggests presets and templates to get started. Or where AI agents collect context autonomously, and emphasize the work, the plan, the tasks — the outcome, instead of the chat input. All of it are examples of great User-First, AI-Second experiences. Not experiences circling around AI features, but experiences that truly amplify value for users by sprinkling a bit of AI in places where it delivers real value to real users. And that’s what makes truly great products — with AI or without. ✤ Useful Design Patterns Catalogs: Shape of AI: Design Patterns, by Emily Campbell 👍 https://shapeof.ai/ AI UX Patterns, by Luke Bennis 👍 https://lnkd.in/dF9AZeKZ Design Patterns For Trust With AI, via Sarah Gold 👍 https://lnkd.in/etZ7mm2Y AI Guidebook Design Patterns, by Google https://lnkd.in/dTAHuZxh ✤ Useful resources: Usable Chat Interfaces to AI Models, by Luke Wroblewski https://lnkd.in/d-Ssb5G7 The Receding Role of AI Chat, by Luke Wroblewski https://lnkd.in/d8xcujMC Agent Management Interface Patterns, by Luke Wroblewski https://lnkd.in/dp2H9-HQ Designing for AI Engineers, by Eve Weinberg https://lnkd.in/dWHstucP #ux #ai #design

15 weeks left before the first rules of the AI Act come into effect. Struggling with where to start on AI implementation and compliance? Start with a multidisciplinary team; conduct an AI inventory; carry out AI Impact Assessments; draft AI policies; amend contracts, policies, and data protection documents to reflect AI’s role in your organisation. Ensure your team is trained in AI literacy, as required under the AI Act. To navigate AI implementation and compliance under the EU AI Act, companies must begin by understanding its scope and risk-based approach. The Act categorises AI systems into prohibited, high-risk, or general-purpose. Prohibited AI systems (the first rules coming in) include those exploiting vulnerabilities or engaging in certain AI emotional recognition. High-risk systems, such as those used in management of critical infrastructure, require strict oversight, including documentation, risk assessments, and ongoing monitoring. General-purpose AI systems, widely used across industries, may also face regulatory scrutiny due to their broad impact. The first step for companies is conducting a comprehensive AI inventory. This involves cataloguing all AI systems in use or under development to determine their classification under the AI Act. Through this inventory, companies can assess their compliance obligations and identify any systems that may need modification or discontinuation to meet the Act’s standards. Data protection is a cornerstone of AI compliance. The AI Act mandates that data used in AI systems be high quality, representative, and free from bias. This is especially crucial for high-risk systems, which must undergo continuous risk assessments to protect fundamental rights. GDPR compliance is also essential for any AI system that processes personal data, and companies must ensure their data governance strategies focus on transparency, accountability, and safeguarding individual rights. Contracts are a critical component of AI implementation. Organisations must revisit and amend contracts to address how AI impacts their legal and operational frameworks. These amendments should explicitly cover liability for AI-generated decisions, intellectual property ownership of AI-generated outputs, and data protection compliance. Contracts must minimise legal exposure. Additionally, intellectual property issues around AI, such as ownership of outputs or the use of third-party data, should be clearly defined in these agreements. Following the AI inventory, companies must conduct an AI impact assessment. This assessment includes both a Data Protection Impact Assessment (DPIA) and a Fundamental Rights Impact Assessment (FRIA). The extraterritorial scope of the AI Act means that even non-EU companies must comply if their AI systems impact the EU market. Non-compliance can result in significant fines, making early compliance essential. 15 weeks left to comply.

A new academic paper "AI Agents Under EU Law" is the most thorough regulatory mapping of agentic AI I've read. If you're deploying AI agents in financial services, here's what it means in practice. The paper's central argument is deceptively simple: the regulatory trigger isn't your AI architecture. It's what your agent actually does - which data it touches, which systems it connects to, whose rights it affects. Classification comes second. The record keeping, or inventory, of external actions comes first. Three findings stood out: 1. Governance must be outside the model A system prompt telling an agent "don't do X" is not a security control. It's a suggestion. Article 15(4) of the AI Act requires restricted actions be enforced at the API level - where the tool interface simply doesn't expose the capability. An agent processing trade data needs read access. Not send permissions. Not delete rights. The law now codifies what good architecture has always required: governance as structure, not policy. 2. Behavioural drift is a compliance trigger Agents that improve through Human-in-the-Loop feedback may be crossing a legal threshold, or "substantial modification", under Article 3(23). Adaptive improvement is a feature. It's also, potentially, a regulatory event... State needs to be versioned and behavioural drift needs to be monitored. 3. DORA is already active Everyone is focused on the AI Act deadlines. But for any agent deployed in financial services, DORA has applied since January 2025. ICT third-party risk management, incident reporting, contractual obligations - these aren't future concerns. The paper concludes that the foundational compliance task isn't classifying your AI system. It's drawing the map - every external action, every data flow, every connected system, every affected person. The compliance clock is already running. Most of the map hasn't been drawn yet. AI Agents under EU Law paper https://lnkd.in/eJF6YrSr Full regulatory timeline in the comments too.

𝐀𝐈 𝐂𝐨𝐦𝐩𝐥𝐢𝐚𝐧𝐜𝐞 & 𝐃𝐚𝐭𝐚 𝐏𝐫𝐨𝐭𝐞𝐜𝐭𝐢𝐨𝐧 𝐋𝐚𝐰𝐬 𝐟𝐨𝐫 𝐆𝐞𝐧𝐀𝐈 𝐀𝐩𝐩𝐬 Building GenAI Apps for a Global Audience?  Understanding Regional Data Protection and AI laws is not optional, it is foundational. Here is what you need to know: 1. UNDERSTANDING GLOBAL REGULATORY VARIANCE Building GenAI for a global audience requires understanding regional data protection and AI laws. Key Regulations by Region: • EU AI Act: Risk-based AI obligations for certain AI systems and transparency use cases • GDPR (EU): Transparency & Consent • DPDP (India): Digital Personal Data Protection • PIPL (China): Strict Data Localization • CCPA (California): Data Access & Opt-Out • LGPD (Brazil): Local Compliance Rules 2. IMPACT OF THESE REGULATIONS ON YOUR AI TRAINING DATA To build compliant GenAI apps,  Ensure that data used for training AI models follows the regional rules: Data Collection → Processing → Model Training → Deployment Three Core Requirements: a. User Consent: Obtain explicit consent for data collection and use b. Data Minimization: Collect only necessary data for the intended purpose c. Anonymization: Remove personally identifiable information from training data 3. MITIGATING AI ETHICS AND BIAS RISKS AI systems must be fair and ethical, particularly in high-risk areas: a. Fairness: Ensure your AI models don't discriminate, especially in areas like recruitment or finance. b. Bias Mitigation: Regularly test and adjust your models to reduce bias in the outputs. 4. ENSURING TRANSPARENCY IN AI MODEL DEVELOPMENT Transparency is a cornerstone of compliance, especially when your AI impacts users directly: a. Explainability: Protect data in transit and at rest. b. Consent Management: Collect, track, and manage user consent. c. Privacy by Design: Embed privacy into every system layer. 5. MANAGING CROSS-BORDER DATA FLOW GenAI apps often rely on data from various regions, so it's critical to understand data sovereignty laws: a. Data Sovereignty: Follow local laws on where data is stored and processed. b. Data Transfer Agreements: Use SCCs or BCRs for compliant cross-border transfers. THE COMPLIANCE CHECKLIST Before launching GenAI globally, verify: 1. Regional Compliance: • GDPR for EU? (Transparency & Consent) • DPDP for India? (Data Protection) • PIPL for China? (Data Localization) • CCPA for California? (Access & Opt-Out) • LGPD for Brazil? (Local Rules) 2. Training Data: • User consent obtained? • Data minimized? • PII anonymized? 3. Ethics & Bias: • Fairness tested? • Bias mitigation in place? 4. Transparency: • Explainability documented? • Consent management system? • Privacy by design? 5. Cross-Border: • Data sovereignty compliance? • Transfer agreements (SCCs/BCRs)? Each region has different requirements.  Build for the strictest, adapt for the rest. Which regulation applies to your GenAI app?

I went to an AI UX workshop last night expecting recycled LinkedIn advice about "building AI trust through transparency." Instead, Isabella Yamin tore down LinkedIn's job posting flow using her CarbonCopies AI framework in real-time, while founders shared raw implementation struggles. It completely changed how I'm rethinking Maibel's onboarding flow. Here's what I stole from B2B SaaS principles to redesign emotional AI for B2C: 1️⃣ Progressive disclosure with purpose LinkedIn's fatal flaw? Optimizing for completion ease > Outcome quality. Recruiters are drowning in irrelevant applications because AI never learns what "qualified" means. The personalization paradox: How do we give users enough control without overwhelming them? Users don't want "frictionless". They want INFORMED control. 📌 At Maibel: I was falling into the same trap, making emotional coaching setup so simple that the AI couldn't understand user context. Now? Progressive complexity with clear trade-offs. Show users how their choices impact outcomes. → Want deeper insights? Add more context. → Want faster setup? Here's what the AI can't personalize. 2️⃣ Closed-loop data intelligence: What Platfio gets right They've built a platform for software agencies where where every data point feeds back into the entire system. User preferences in marketing flows shape proposals. Campaign performance shapes future recommendations. Every interaction becomes intelligence for future recommendations. 📌 At Maibel: Most wellness apps store emotional check-ins like digital journals. I'm turning them into predictive feedback loops. Emotional intelligence isn’t static but COMPOUNDS. Today's reflections shift tomorrow's suggestions. Patterns fuel prevention. Users' inputs on Monday could predict AND prevent Friday's breakdown. 3️⃣  Multi-modal creativity: Wubble's transparency approach Translating images and files into music - who'd have thought? They've cracked multi-modal creativity where users become co-creators, not passive consumers. The breakthrough moment for me: What if users could see how their visual environment contributes to emotional context? 📌 At Maibel: Users upload images of their day and see how AI analyzes emotional cues: cluttered workspace = overwhelm, junk food = stress eating. Multi-modal understanding users can contribute to and influence. 💡 The bottom line? B2B Saas gets one thing right: Every interaction has to earn trust. In B2B, failed AI means churn. In emotional AI, failed trust breaks belief in tech entirely. 📌 Here's what we're doing differently at Maibel: → Progressive complexity → Context-aware feedback → Multi-modal participation → Intelligence that compounds with every input. It's not just about building WITH AI. I'm designing systems that learn understand YOU before you even need to explain yourself. Kudos to Isabella, Shivang Gupta The Generative Beings, Shaad Sufi Hayden Cassar and everyone who shared deep product insights.

Giving users clear insight into how AI systems think is a smart business strategy that builds loyalty, reduces friction, and keeps people from feeling like they’re at the mercy of a mysterious black box. Explainable AI (XAI) enhances the transparency of AI decision-making, which is vital for customer trust—especially in sectors like finance or healthcare, where stakes are high. Tools like SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-Agnostic Explanations) break down complex algorithms into interpretable outputs, helping users understand not just the “what” but the “why” behind decisions. Interactive dashboards translate this data into visual forms that are easier to digest, while personalized explanations align AI insights with individual user needs, reducing confusion and resistance. This approach supports more responsible deployment of AI and encourages wider adoption across industries. #AI #ExplainableAI #XAI #ArtificialIntelligence #DigitalTransformation #EthicalAI

AI explainability is critical for trust and accountability in AI systems. The report “AI Explainability in Practice” highlights key principles and practical steps to ensure AI decisions are transparent, fair, and understandable to diverse stakeholders. Key takeaways: • Explanations in AI can be process-based (how the system was designed and governed) or outcome-based (why a specific decision was made). Both are essential for trust. • Clear, accessible explanations should be tailored to stakeholders’ needs, including non-technical audiences and vulnerable groups such as children. • Transparency and accountability require documenting data sources, model selection, testing, and risk assessments to demonstrate fairness and safety. • Effective AI explainability includes providing rationale, responsibility, safety, fairness, data, and impact explanations. • Use interpretable models where possible, and when black-box models are necessary, supplement with interpretability tools to explain decisions at both local and global levels. • Implementers should be trained to understand AI limitations and risks and to communicate AI-assisted decisions responsibly. • For AI systems involving children, additional care is required for transparent, age-appropriate explanations and protecting their rights throughout the AI lifecycle. This framework helps organizations design and deploy AI that stakeholders can trust and engage with meaningfully. #AIExplainability #ResponsibleAI #HealthcareInnovation Peter Slattery, PhD The Alan Turing Institute

"A Multifaceted Vision of the Human-AI Collaboration: A Comprehensive Review" provides some interesting and useful insights into effective Humans + AI work, drawn from across the literature. Some of the specifics insights in the paper: 🧭 Use the five-cluster framework to tailor collaboration depth. The framework defines five types of human-AI collaboration: (1) Humans as optional tools, (2) Consensus-based coordination, (3) Asynchronous collaboration, (4) Humans and AI as co-agents, and (5) Humans directing AI. Choose the type based on your task: use cluster 1 for personalization (e.g. recommender systems), cluster 2 for group decision-making, clusters 3 and 4 for task co-execution, and cluster 5 when human judgment must lead the process. 🧠 Let humans steer the learning loop. Design workflows where human feedback isn't just collected but actively changes the model. Show users how their input influences outcomes, and ensure systems update based on their corrections—failing to do so erodes trust and engagement fast. 🔄 Support iterative improvement through clear feedback cycles. Let users provide input at multiple points in the workflow—before, during, and after AI output. Use real-time feedback, editable suggestions, and memory-based personalization (e.g., saving past preferences) to refine collaboration with each loop. 📣 Grant users communication initiative. Don’t restrict user interaction to predefined prompts—enable them to ask questions, challenge decisions, or suggest new directions. This increases user autonomy, supports trust, and improves performance in both individual and group collaboration. 🛠️ Customize AI outputs to user-specific contexts. Embed features that allow tailoring of recommendations, predictions, or decisions to individual preferences or needs. For example, let users tweak rehabilitation goals in health tools or input content preferences in recommender systems. 🤖 Use AI as an impartial coordinator in group settings. In scenarios with multiple human participants—such as disaster planning or multi-user workflows—deploy AI to synthesize input, allocate tasks, and reduce bias. Ensure the system is transparent and users can reject or adjust AI decisions. 🔐 Prioritize human-centered design values. Build systems that are transparent (explain why outputs were generated), trustworthy (learn from user feedback), accessible (usable by non-experts), and empowering (give users control over high-level behavior). These are essential for lasting, ethical collaboration.