<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Devaki Joshi on Medium]]></title>
        <description><![CDATA[Stories by Devaki Joshi on Medium]]></description>
        <link>https://medium.com/@devkii_?source=rss-531ea0f2de86------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*pnZHBs0NAMQIAqtm83Mraw.jpeg</url>
            <title>Stories by Devaki Joshi on Medium</title>
            <link>https://medium.com/@devkii_?source=rss-531ea0f2de86------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 14:32:37 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@devkii_/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Digital Handshakes: The Strategic Design of Modern Payments]]></title>
            <link>https://medium.com/@devkii_/the-invisible-wall-balancing-ironclad-security-with-a-frictionless-checkout-10b54c306e96?source=rss-531ea0f2de86------2</link>
            <guid isPermaLink="false">https://medium.com/p/10b54c306e96</guid>
            <category><![CDATA[design-thinking]]></category>
            <category><![CDATA[payment-gateway]]></category>
            <category><![CDATA[security]]></category>
            <category><![CDATA[ui-design]]></category>
            <category><![CDATA[ux-design]]></category>
            <dc:creator><![CDATA[Devaki Joshi]]></dc:creator>
            <pubDate>Sun, 15 Feb 2026 12:32:24 GMT</pubDate>
            <atom:updated>2026-02-16T16:35:35.048Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*G8JIpUQI6HUmfm_5L21Xxg.png" /></figure><p>I remember the very first time I made an online transaction for a Sport’s Jacket on Amazon. Making the payment was almost walking on eggshells, feeling my heart race with every loading icon. I was rechecking every little detail, going back and forth and obsessively verifying the URL every thirty seconds to ensure I hadn’t wandered onto a phishing site. Fast forward 2 months; I was shopping online at least thrice a week, compulsively hitting buy now every chance I’d get. The once-daunting ‘Buy Now’ button was now a rhythmic, almost mindless part of my routine.</p><p>Being a design enthusiast, it made me wonder, for designers, it’s hard enough to retain users, and even harder to make them shell out their hard-earned money. How did these platforms manage to gain enough trust to be such a widely used service?</p><p>After I deep-dived in, I realized this seemingly short 1 min process is backed by extensive research, complex architecture of secure user flow and systematic execution of data-driven design. It’s a delicate balance of minimizing cognitive load while maintaining the rigorous encryption protocols necessary for a secure payment gateway.</p><p>Payment gateways are crucial as they are ‘make or break’ points of a user’s journey. If the gateway takes half a second too long or at any point the users feel their data is not secure, they will quit the platform and most probably never return again. Hence, better design is not about cosmetics, it’s revenue.</p><p>The payment gateway is the digital interaction between buyer and merchant and is divided into 3 stages: data entry (the eggshell moment where you enter your details), data verification (the guard ensures data accuracy and checks for scams) and finally authorization (the ultimate handshake that seals the deal).</p><p>From a UI/UX perspective, data entry stage is where users might experience most friction (anxiety) as they face high-stakes pressure of inputting sensitive financial information correctly as well as fear of security breaches while navigating through and figuring out the flow of the platform. A clean, transparent, user-friendly interface during this phase followed by a clear data-driven verification and authorization is where we can quietly aid users complete their transaction.</p><p>Now, is that enough?</p><p>Definitely not! Because no matter how clean and perfect your design is, something can always go wrong. Maybe you enter the wrong UPI address, or midway realize the account number you enter has insufficient balance, or the internet disconnects mid payment. Payment failures generally fall into two categories: Soft Declines, which are temporary technical hiccups like network timeouts or gateway glitches; and Hard Declines, which are definitive rejections due to permanent issues like insufficient funds or incorrect credentials.</p><p>Handling these issues isn’t just a backend headache, it also requires an empathetic UI that provides clear, actionable feedback as well as graceful error handling and intuitive recovery paths. These prevent technical hiccups and guide the user back to safety without making them feel like they’ve failed the process.</p><p>Here are 5 such must do’s to keep in mind while designing a seamless, payment gateway:</p><ol><li><strong>Utilize familiar visual cues: the 3-second rule</strong></li></ol><p>The decision of whether or not the user trusts the page is decided in the first 3 seconds. Thus, before a user even types a single digit, they scan the page for safety.</p><p><em>The Design: </em>Displaying SSL padlocks, PCI-DSS badges, and recognized card logos (Visa, Mastercard, UPI).</p><p><em>The Impact: </em>Gains user trust and promises transparency.</p><p><strong>PRO-TIP:</strong> Avoid too much clutter here. This creates distractions, resulting in increased confusions. Keep concise, minimalistic UI’s.</p><p><strong>2. Provide diverse payment options</strong></p><p>Now that they trust you, they look for their preferred way to pay.</p><p><strong><em>The Design: </em></strong>Offer a mix of Credit/Debit, UPI, Wallets, and Net Banking.</p><p><strong><em>The Impact: </em></strong>It caters to different user habits — whether it’s the compulsive shopper using a one-tap UPI intent or a cautious buyer using a credit card for extra protection.</p><p><strong>PRO-TIP: </strong>Keep the selecting payment option and entering details on separate pages, with a proceed to pay option in between. This prepares the user to keep their card details/credentials ready. You can also add a small micro-copy hint like “You’ll need your card number and CVV for the next step”.</p><p><strong>3. Thumb-First Rule</strong></p><p><strong><em>The Design:</em></strong></p><p>(i) <strong>The Green Zone:</strong> Place all critical interactive elements — the Proceed to Pay button and Payment Method toggles — within the bottom 1/3 to 2/3 of the screen.</p><p>(ii) <strong>Generous Targets: </strong>Ensure buttons are at least 44x44 pixels to prevent fat-finger errors.</p><p>(iii) <strong>Bottom-Anchored CTAs: </strong>Stick the primary action button to the bottom of the viewport so it stays accessible regardless of page length.</p><p><strong><em>The Impact: </em></strong>The UI becomes comfortable and lighter, reducing physical strain and reducing psychological barriers.</p><p><strong>PRO-TIP: </strong>This works best for UPI/mobile payments.</p><p><strong>4. Real-time feedback and Micro-Validations</strong></p><p><strong><em>The Design: </em></strong>Use real-time feedback (Micro-Validations) to catch typos as they happen.</p><p>(i) Use instant visual cues like green checkmarks or subtle red borders as the user finishes typing in each field.</p><p>(ii) Provide specific, helpful prompts like <em>“Check your expiry date”</em> rather than generic messages.</p><p><strong><em>The Impact: </em></strong>This is a major key to increase user trust and confidence.</p><p>(i) Cognitive Ease: Users correct mistakes in the moment, preventing the frustration of a “page refresh” or a massive list of errors at the end.</p><p>(ii) Boosted Confidence: Seeing a field turn green provides a “micro-win,” signaling to the user that they are on the right track and the system is secure.</p><p>(iii) Zero-Friction Checkout: Minimizes form abandonment by ensuring the final “Pay” click is a guaranteed success rather than a gamble.</p><p><strong>5. Clear Communication and Flow: Progress Bars</strong></p><p><strong><em>The Design:</em></strong></p><p>(i) Step-by-Step Visualization: Use a linear progress bar at the top of the screen labeled with clear stages (e.g., Method → Details → Confirm).</p><p>(ii) Active States: Highlight the current step with a distinct color or bold icon, while graying out pending steps to maintain focus.</p><p>(iii) Breadcrumb Logic: Allow users to click back to previous steps (eg: Method Selection) directly from the bar if they change their mind.</p><p><strong><em>The Impact: </em></strong>This major design change eliminates uncertainty, thus minimizing the risk of cart abandonment.</p><p><strong>6. Transparent Loading and Status States: The Processing Wait</strong></p><p><strong><em>The Design:</em></strong></p><p>(i) Contextual Messaging: Replace generic spinners with status-driven text like “Verifying with your bank…” or “Verifying your transaction…”</p><p>(ii) Visual Feedback: Use a branded loading animation or a progress ring that fills to show the system is actively working, not frozen.</p><p>(iii) The “No-Go” Shield: Programmatically disable the “Pay” button and “Back” gesture once clicked to prevent double-charging or interrupted sessions.</p><p><strong><em>The Impact:</em></strong></p><p>(i) Patience Engineering: Informative labels make a 3-second wait feel shorter and more purposeful than a blank screen or a static icon.</p><p>(ii) System Trust: Transparency during the highest-stakes moment (the actual charge) reassures the user that their money is being handled safely.</p><p>(iii) Technical Integrity: Prevents duplicate orders and server-side errors by clearly signaling that the request is already in flight.</p><p><strong>7. Empathetic Error Handling: The “Unhappy Path”</strong></p><p><strong><em>The Design:</em></strong></p><p>(i) Human-Centric Copy: Swap robotic codes (e.g., “Error 402”) for supportive, plain language like “Your card was declined — double-check the CVV or try a different method.”</p><p>(ii) Persistent Field Data: Retain all correctly entered information so the user only has to fix the specific field that failed.</p><p>(iii) Actionable Fixes: Provide a direct “Try Again” or “Switch Payment Method” button immediately below the error message.</p><p><strong><em>The Impact:</em></strong></p><p>(i) Softens the shock of a failed transaction, preventing the user from feeling blamed or anxious about their finances.</p><p>(ii) By saving valid data, you eliminate the start-over fatigue that usually leads to immediate cart abandonment.</p><p>(iii) Transforms a potential dead-end into a guided detour, keeping the user in the funnel until the purchase is successful.</p><p>Your payment flow is more than just a transactional endpoint — it’s a critical moment of trust, clarity, and conversion. Done poorly, it erodes confidence and pushes users away. Done right, it builds loyalty, increases lifetime value, and becomes a competitive edge.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=10b54c306e96" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beyond the Algorithm: Technical Deep-Dive at the ACM Winter School]]></title>
            <link>https://medium.com/@devkii_/beyond-the-algorithm-my-technical-deep-dive-at-the-acm-winter-school-9d12c8cb2a29?source=rss-531ea0f2de86------2</link>
            <guid isPermaLink="false">https://medium.com/p/9d12c8cb2a29</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[responsible-ai]]></category>
            <category><![CDATA[health-it-trends]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[neural-networks]]></category>
            <dc:creator><![CDATA[Devaki Joshi]]></dc:creator>
            <pubDate>Thu, 15 Jan 2026 08:58:51 GMT</pubDate>
            <atom:updated>2026-01-15T09:11:47.215Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zjxQZwL0HWeuG5R_oIIeZw.jpeg" /></figure><p>You’ve spent years building a career in tech. You apply for a job at a major corporation. You’re highly qualified, but your resume is rejected by the AI in milliseconds because you mentioned you were the president of the “Women’s Chess Club”. Many companies, including major MNC’s had to scrap a secret AI recruiting tool after realizing it was actively penalizing resumes that included the word “women’s”. The AI was trained on 10 years of successful resumes, most of which came from men. It learned that being “male” was a success factor and started filtering out anyone who didn’t fit that historical mold.</p><p>Whenever building any ML models, we usually always have this one question, <em>“How do we make this model 1% more accurate?”. </em>But my recent experience of ACM Winter School on Responsible AI, held in XIM University changed my approach towards AI and raised an entirely new question in my mind.<em> “If a model is perfectly accurate based on historical data, but that data is biased, is the model ‘intelligent’ or is it just a high-speed mirror for our own prejudices?”</em></p><p>Held by top professors and researchers across the country, I had the opportunity of deep-diving into the mechanics of Responsible AI, Network Science, and Graph Neural Networks (GNNs) at the ACM Winter School.</p><p>My goal was to move beyond the “black box” of technology and understand how we can teach these high-speed machines to follow human values instead of just mimicking historical mistakes. Here is what I learned about building a more equitable digital future.</p><p><strong>I. Algorithmic Accuracy vs. Fairness</strong></p><p>In the opening days, we understood Fair Classification and Recommendations, including their importance in making critical decisions, applications in the real-world and challenges encountered in building them.</p><p>One may think that if we simply remove a person’s gender or race (called sensitive attributes) from a dataset, the AI cannot discriminate. However, we learnt that algorithms are experts at finding “proxies” — hidden patterns like zip codes or shopping habits that secretly give away those protected traits.</p><p>ML algorithms tend to make decisions that align with their trained data, this tendency is called Empirical Risk Minimization (ERM). An algorithm treats historical data as ground truth and optimizes a loss function to minimize the average error across the provided dataset.</p><p>Moreover, algorithms cannot distinguish between structural causal relationships and societal noise. When biased model outputs are used to make real-world decisions, they generate new data points that reinforce the original bias. This creates a self-fulfilling prophecy where the training data for the next generation of models is even more skewed.</p><p>Even when protected attributes (like race or gender) are excluded, the model exploits proxy variables — features that are highly correlated with the excluded attributes — to maintain predictive performance. This results in disparate impact, where the model’s decision boundary implicitly discriminates against specific subgroups.</p><p>While disparate impact occurs when an apparently neutral process results in biased outcomes, we must also consider the more direct violation of fairness: disparate treatment.</p><p>While disparate treatment is an explicit violation — occurring when a model’s prediction Y is directly conditioned on a protected attribute A, disparate impact is often more insidious because it is implicit.</p><p>During our sessions, we learned that addressing these two issues requires different technical interventions:</p><ul><li><strong>To solve Disparate Treatment:</strong> we use Anti-classification, ensuring A is removed from the feature set and the model is “blind” to the sensitive attribute.</li><li><strong>To solve Disparate Impact: </strong>we must move beyond blindness and use Independence constraints, forcing the model to ensure that the distribution of Y is independent of A.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JFvfRzd5ufFe_bbQf81ZpQ.png" /></figure><p>We further went on to study how to reduce discrimination in ML algorithms.</p><ul><li><strong>Pre-processing:</strong> This involves transforming the training data before the model is built to remove underlying biases and correlations with protected attributes. Techniques such as re-weighing, oversampling, and label massaging to neutralize underlying biases and correlations.</li><li><strong>In-processing:</strong> This integrates fairness constraints directly into the model’s optimization process during training. We use adversarial debiasing and add fairness penalty terms to the objective function to force the algorithm to prioritize equity.</li><li><strong>Post-processing:</strong> This adjusts the model’s output scores or classification thresholds after training to ensure equitable final results. This stage recalibrates the model’s final scores after training is complete, applying techniques like equalized odds thresholding or calibrated odds adjustment to ensure the final decisions meet fairness criteria.</li></ul><p>Fair recommendations extend the concept of equity beyond individual classification to the dynamic selection of content in information-retrieval systems. In these environments, the objective is to mitigate feedback loops where algorithms favor historically popular items or majority demographics, which systematically suppresses the visibility of minority groups or new entrants.</p><p>This was followed by a hands-on session focused on fairness in food delivery platforms, a complex three-sided market involving customers, restaurants, and delivery partners. We explored how bias can manifest in order allocation and covered long-term fairness concepts such as amortized parity (aiming for equal outcomes over time), amortized proportionality, and the egalitarian objective (maximizing the utility of the worst-off participant).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tGZv84WUsW6f3bFlj4usoA.png" /></figure><p>To implement these concepts practically, we followed a structured algorithmic approach:</p><ul><li><strong>Step 1:</strong> Assign edge weights to represent the cost or priority of a delivery assignment.</li><li><strong>Step 2:</strong> Identify all nearby drivers within a radius where the restaurant-to-driver distance ensures the Estimated Delivery Time (EDT) does not exceed the Service Level Agreement (SLA).</li><li><strong>Step 3: </strong>Construct a bipartite graph connecting available orders to eligible drivers.</li><li><strong>Step 4:</strong> Execute a minimum cost matching algorithm on the graph to optimize for both efficiency and the defined fairness constraints.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/834/1*LG5JMMzOpFieg0bqXbCvBQ.png" /></figure><p><strong>II. The Topology of Thought: Understanding Brain Network Analysis</strong></p><p>Moving from the ethical constraints of AI, the winter school shifted focus toward Network Science. We explored how complex systems — ranging from the biological neurons in our brains to the social ties in our communities — can be modeled as mathematical graphs.</p><p>We started by learning about the connectome, a matrix representing all possible pairwise anatomical connections between neural elements of the brain. We explored how modeling the brain as a graph allows us to analyze functional and structural connectivity through the lens of graph theory.</p><p>Our exploration traced the evolution of network theory through several foundational concepts:</p><ol><li><strong>Seven Bridges of Königsberg:</strong> Euler’s 1736 solution established the foundations of Graph Theory, demonstrating that system traversability is a function of node degrees and topological connectivity rather than Euclidean distance.</li><li><strong>Small-World Phenomenon:</strong> Characterized by the Watts-Strogatz model, these networks exhibit high local clustering and low average path lengths, facilitating efficient global information transfer (the mathematical basis for Milgram’s 1967 Experiment and Bacon’s Oracle).</li><li><strong>Barabási-Albert Model (1999):</strong> This model describes Scale-Free networks driven by preferential attachment, where the probability of a node receiving new edges is proportional to its current degree, resulting in a power-law distribution.</li><li><strong>Pareto Principle:</strong> A consequence of scale-free dynamics where approximately 80% of network traffic is handled by 20% of the nodes. In neurobiology, these “hubs” form rich-club architectures critical for high-level cognitive integration.</li></ol><p>We define a neuron as the fundamental structural and functional unit of the brain. Building from this unit, we explored two ways these neurons interact: Structural Connectivity, the physical “hardware” of anatomical fiber bundles and synapses, and Functional Connectivity, the “software” representing statistical patterns of communication between brain regions.</p><p>We examined how network science is applied across three distinct biological scales: microscopic, mesoscopic and macroscopic.</p><ul><li><strong>Microscopic:</strong> networks reconstructed at the level of individual neurons and synapses. This requires electron microscopy to map the precise structural connections of local circuits.</li><li><strong>Mesoscopic:</strong> Analyses at this scale combine microscopic and macroscopic techniques to understand neuronal connectivity with high precision across the entire brain, with a goal to understand neuronal connectivity with high precision across larger brain regions or the entire brain. This is done using viral tracing or light-sheet microscopy.</li><li><strong>Macroscopic: </strong>This level focuses on structural and functional interactions between large-scale populations of neurons. It utilizes non-invasive techniques such as MRI (DTI) for structural fiber tracts and MEG or EEG for real-time functional dynamics.</li></ul><p>We then further understood the neuron theory with Ramón y Cajal’s silver impregnation experiment. Ramón y Cajal claimed that neurons were discrete cells that contacted each other very closely by synaptic junctions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*AilPzYl5vYkB_cNMF_1rXA.png" /><figcaption>Fig: MRI image at different planes</figcaption></figure><p>Nodes can be either individual neurons (micro), groups of neurons (meso) or voxels(macro) axonal projections or synaptic junctions can be represented by edges.</p><p>After understanding the basics of diffusion MRI as well as EEG, MEG, we went on to discuss functional MRI. EEG and MEG have poor spatial resolution that poses special problems. The diffuse signal recorded by each sensor, which aggregates activity across large populations of neurons, makes it difficult to localize effects with precision. Functional MRI generally offers a higher spatial resolution than EEG and MEG, by examining the fluctuations in blood-oxygenation-level-dependent (BOLD) signals.</p><p>To translate these network theories into clinical practice, we analyzed the standard fMRI workflow used to extract functional connectivity data. The process begins with measuring brain activity at the voxel level — the 3D equivalent of a pixel — and follows a rigorous preprocessing pipeline to ensure signal integrity:</p><ol><li><strong>Volume Dropping:</strong> Discarding initial scans to allow for T1-equilibration.</li><li><strong>Distortion &amp; Slice Timing Correction:</strong> Adjusting for magnetic field inaccuracies and the slight time offsets between acquired brain slices.</li><li><strong>Motion Correction:</strong> Realigning images to compensate for the subject’s head movement.</li><li><strong>Spatial Normalization:</strong> Mapping the individual brain to a standard template (like the MNI space) for group comparison.</li><li><strong>Spatial Smoothing:</strong> Applying a Gaussian filter to increase the signal-to-noise ratio and account for anatomical variability.</li></ol><p>To conclude our session on neurobiology, we explored a clinical application of these concepts: Diagnosing ADHD (Attention Deficit Hyperactivity Disorder).</p><p>We studied:</p><ul><li><strong>Graph Representations:</strong> How fMRI-derived data is converted into brain graphs, where nodes represent specialized regions and edges represent the strength of functional synchronization.</li><li><strong>Topological Biomarkers:</strong> Identifying ADHD through shifts in degree distribution and centrality metrics, which reveal that critical communication hubs in executive control networks are often bypassed or weakened.</li><li><strong>Automated Classification:</strong> The use of Graph Neural Networks (GNNs) to ingest these high-dimensional graph structures, allowing for the objective detection of dysconnectivity patterns that are invisible to traditional clinical observation.</li></ul><p>This objective, network-driven approach to ADHD diagnosis demonstrates the profound potential of Graph Neural Networks to transform medical care from reactive treatment to proactive, data-informed intervention.</p><p><strong>III. Holding the Code Accountable: The Science of Algorithmic Auditing</strong></p><p>This session addressed a critical question: <strong>How do we ensure that platforms actually follow the fairness policies they claim to uphold? </strong>We explored how to treat algorithms not as “black boxes,” but as systems that must be systematically inspected to ensure they meet legal and societal standards for fairness and accountability.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eZQ_9wH10oyC0KBPRkQ-ZQ.png" /></figure><p>We first explored accountability — the obligation of platform developers to explain, justify, and take responsibility for the consequences of their algorithmic decisions. We all have a digital presence and our personal data is used to develop AI models that mediate our interactions. Hence, the need for regulations comes into the picture.</p><p>We need regulations to safeguard our privacy, dignity, and autonomy in the digital environment by giving individuals meaningful control over their data, protecting us from data-driven discrimination, fraud, and manipulation, and fostering the trust required for a safe online space.</p><p>We studied the key regulations designed to improve platform accountability:</p><ol><li><strong>GDPR (General Data Protection Regulation):</strong> Focuses on data privacy and the right to an explanation for automated decisions.</li><li><strong>Digital Services Act (DSA):</strong> Targets online safety and transparency in content moderation and recommender systems.</li><li><strong>Digital Markets Act (DMA):</strong> Aims to ensure contestability and fairness by regulating “gatekeeper” platforms.</li><li><strong>AI Act:</strong> A risk-based framework (then in its finalization stages) to regulate AI systems based on their potential for societal harm.</li></ol><p><strong>GDPR : Key definitions, stakeholders and principles</strong></p><p>General Data Protection Regulation (GDPR) is a key regulation established to improve platform accountability and protect the fundamental rights of individuals in the digital economy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UeQvdYGKmFh7MkDUNc_jvA.png" /></figure><p><strong>DSA (Digital Services Act):</strong></p><p>The Digital Services Act (DSA) is a landmark regulation designed to harmonize rules for online intermediaries and ensure a safe, predictable, and accountable online environment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DrXIgcVYaog3TlDQjUVvhA.png" /></figure><p>While studying GDPR and its policies (Right to Access), we stumbled upon Data Download Packages (DDPs) — comprehensive archives of their personal information and activity. In algorithmic auditing.</p><p>We went hands-on and compared DDP’s of Tiktok and GDPR, two very known online social media platforms. While both platforms provide these archives, our comparative analysis revealed significant differences in <strong>reliability</strong>, <strong>completeness</strong>, and <strong>transparency</strong>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/609/1*ATfjLIXekHWiBdT4ECyEPw.png" /></figure><h4>Decentralized Accountability: Transparency on Bluesky</h4><p>Before wrapping up, we explored <strong>Bluesky</strong>, an autonomous platform built on open protocols that challenges the “black box” nature of centralized social media.</p><p><strong>The Problem:</strong> Current transparency databases are hindered by opaque reporting and “flimsy” moderation, where the most prominent violations are often vaguely categorized as “other violations of terms and conditions”.</p><p><strong>The Bluesky Solution:</strong> By decentralizing power and utilizing an open protocol, Bluesky provides users with greater agency and autonomy over their digital experience. Its architecture allows for verifiable, multi-layered moderation rather than a single, hidden authority.</p><p><strong>Core Technical Features:</strong></p><ul><li><strong>Personal Data Servers (PDS):</strong> Users store their own repositories — every action they have ever performed — on independent servers.</li><li><strong>Relays &amp; The Firehose:</strong> Relays crawl these PDSes to create a “firehose,” an aggregated, real-time stream of all platform updates.</li><li><strong>Labelers:</strong> Independent services that consume the firehose to flag and moderate content according to specific community standards.</li><li><strong>Feed Generators:</strong> Specialized algorithms that reorder the firehose to create custom user feeds.</li><li><strong>App View:</strong> The final interface that consumes the processed records and enforces chosen moderation controls for the user.</li></ul><p>As we move from policy to practice, these tools do more than just organize data — they safeguard our digital rights in an increasingly connected world.</p><p><strong>IV: The Secure LLM: Bridging Transformer Architecture with Privacy-First Tuning</strong></p><p>In this session, we explored the end-to-end lifecycle of Responsible LLM Engineering, moving from core structural design to the technical implementation of privacy safeguards. We studied how to build high-performance models that are not only accurate but also respect user confidentiality.</p><p>This session dismantled the “black box” of Large Language Models (LLMs) to understand their structural foundations and the engineering required to make them privacy-aware.</p><h4>1. Structural Foundations: The Transformer Architecture</h4><p>We analyzed the fundamental “hardware” of the LLM, focusing on how data transforms into thought through several key layers:</p><ul><li><strong>The Transformer Core:</strong> Uses self-attention mechanisms and feedforward layers to process complex, long-range dependencies in text simultaneously.</li><li><strong>Input Pipeline:</strong> Processes raw text through Tokenization (breaking text into units), Token Embedding (converting units into vectors), and Positional Encoding (adding sequence order).</li></ul><p><strong>Architectural Variations:</strong></p><ul><li><strong>Encoder-Only (Autoencoders):</strong> Designed for context understanding and extraction (e.g., BERT).</li><li><strong>Decoder-Only (Autoregressors):</strong> Optimized for generative tasks and predicting the next token (e.g., GPT).</li><li><strong>Encoder-Decoder:</strong> Hybrid models used for complex sequence-to-sequence tasks like translation.</li></ul><p>Once the architecture is established, general-purpose models must be refined for specific tasks. We studied three primary methods to align LLMs with specialized requirements:</p><ul><li><strong>Supervised Fine-Tuning (SFT):</strong> Training the model on high-quality, labeled prompt-response pairs to teach it specific behaviors or domain knowledge.</li><li><strong>Parameter-Efficient Fine-Tuning (PEFT):</strong> Methods like LoRA (Low-Rank Adaptation) that freeze the original model weights and only train a tiny fraction of new parameters, significantly reducing memory and compute costs.</li><li><strong>RLHF (Reinforcement Learning from Human Feedback):</strong> Aligning model outputs with human values and safety by using a reward model to “rank” better responses.</li></ul><p>Beyond traditional fine-tuning, we explored <strong>RAG</strong>, a framework that enhances LLM accuracy by allowing the model to look up external information before generating a response.</p><p>We studied its core mechanism:</p><ul><li><strong>The Retrieval Step:</strong> Instead of relying solely on internal weights, the system queries a <strong>vector database</strong> of private or up-to-date documents.</li><li><strong>Contextual Augmentation:</strong> The relevant “chunks” of data are retrieved and fed into the prompt as a reference.</li><li><strong>Fact-Based Generation:</strong> The LLM uses this retrieved context to produce an answer, significantly reducing “hallucinations”.</li></ul><p><strong>The Key Advantage:</strong> While fine-tuning teaches a model new skills, RAG provides it with a reliable knowledge base that can be updated instantly without retraining.</p><p>In conclusion, we addressed some key open challenges in LLM training:</p><ol><li><strong>Memorization vs Generalization</strong> : How can we reduce memorization without hurting performance?</li><li><strong>Catastrophic Forgetting</strong> : While fine-tuning LLMs, they tend to forget prior knowledge.</li><li><strong>Privacy breach :</strong> LLMs in its process of memorization may inadvertently spill out PII.</li><li><strong>Context-window limitations</strong> : Larger the lookback range, more the tendency to lose attention.</li></ol><p><strong>V. Decoding Smart Healthcare: The Pulse of Explainable AI</strong></p><p>Imagine a patient, let’s say John, who visits the ER with vague chest pains. However, John has to wait hours in crowded hallways just to fill the application form, due to resource constraints. When he is finally seen, the approach is reactive: doctors treat the symptoms he has <em>now</em>, rather than having caught the subtle heart rate changes that started days ago on his smartwatch.</p><p>The current healthcare facility is resource constrained, hospital-centric and reactive in nature. There are other challenges such as high data-collection and pre-processing costs.</p><p>Thus, a new, proactive approach towards making healthcare more accessible, efficient and cost-effective is Smart Healthcare. It assures point of care diagnosis and remote monitoring telehealth capabilities.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*03cuayVqXfe9VKtwxQuz0g.png" /></figure><p>We explored the three core technical frameworks of smart healthcare:</p><ol><li><strong>Internet of Medical Things (IoMT):</strong> A connected infrastructure of medical devices and software that gathers and transmits real-time physiological data to the cloud for continuous monitoring.</li><li><strong>Digital Health (DT):</strong> The broad use of information and communication technologies — ranging from mobile health apps to electronic records — to manage health risks and personalize the patient journey.</li><li><strong>Healthcare Cyber-Physical Systems (H-CPS):</strong> The “brain” of the operation, where computational algorithms and physical medical processes work in a closed loop. H-CPS doesn’t just monitor; it uses real-time feedback to trigger automated responses, such as alerting a clinic before a cardiac event occurs.</li></ol><p>IoMT deals with data collection, data transmission, and data processing. It is a connected infrastructure of medical devices and software that gathers and transmits real-time physiological data to the cloud for continuous monitoring.</p><p>We then learnt different data mining and feature extraction methods involved in Smart Healthcare, during which we were also introduced to TinyML.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gBDb0p311wu8ZnDveY8bjQ.png" /></figure><p>The transition to Smart Healthcare is ultimately a masterclass in managing technical trade-offs. While high-fidelity deep learning models offer high computational efficiency and diagnostic accuracy, it is difficult to run them on the low-power microcontrollers of IoMT devices. This is a problem that TinyML solves. While it sacrifices some model complexity (via quantization and pruning) to gain real-time, on-device intelligence that functions without cloud latency.</p><p><strong>VI. Securing the Spark: Privacy-First AI</strong></p><p>In today’s world, ML models contain more of our personal data than ever before. We shop online, look for our favorite nearby cafes, have our smart watches on, and consume tons of social media content. Thus, everything ranging from where we are to at what rate our heart is beating is harvested to train an increasingly sophisticated algorithm. These ML algorithms ultimately are used as products that impact us in day to day lives, hence their privacy is a massive liability. This session addressed this liability and introduced us to what is meant by ‘Privacy of AI’.</p><p>Privacy attacks on machine learning models can compromise confidentiality by leaking model parameters and training data. Any unauthorized access to this information is classified as a data breach. Privacy attacks can be from a white box attacker (when an attacker knows the model architecture, parameters etc) or a black box attacker (when the attacker doesn’t).</p><p>Common privacy attacks include membership inference attacks, where an attacker determines whether a specific data point was used during training, and model inversion attacks, which attempt to reconstruct sensitive attributes or training samples. In some cases, model extraction attacks can also reveal model parameters, indirectly exposing patterns present in the training data. Overfitting increases vulnerability, as models that memorize training data are more likely to leak information, and smaller groups are potentially more vulnerable.</p><p>We then went on to study privacy-preserving techniques such as differential privacy, data anonymization, and secure model access controls to limit information exposure while maintaining model utility.</p><p>We also studied cryptographic approaches to mitigate privacy risks such as homomorphic encryption and secure multi-party computation. These involve performing computations on encrypted data so that sensitive information remains protected even during computation. It supports additive and multiplicative operations on cipher text.</p><p>As ML systems become more deeply embedded in everyday applications, designing models that are not only accurate but also privacy-aware is essential. Ensuring the privacy of AI systems is no longer optional — it is a critical requirement for building trustworthy, ethical, and responsible machine learning solutions.</p><p><strong>VII. The Shape of Thought: Mapping Knowledge with Graph Neural Networks</strong></p><p>Most traditional AI models, like CNNs for images, view the world as a rigid grid. But the most interesting parts of our lives — our social circles, the chemical bonds in a new medicine, and even the functional pathways in our brains — don’t fit into neat boxes. They are graphs. This session explored Graph Neural Networks (GNNs), a specialized class of deep learning designed to reason about these complex, “non-Euclidean” relationships where meaning comes from connection, not just positions.</p><p>We differentiated between homogeneous graphs, which consist of a single type of node and edge, and heterogeneous systems, which involve multiple entity types, such as those found in medical databases. A key focus was the contrast between homophily, where connected nodes tend to be similar, and heterophily, where connected nodes are dissimilar. In heterophilous settings, standard aggregation methods can cause excessive signal smoothing, making specialized aggregators necessary to preserve meaningful distinctions and ensure predictive accuracy.</p><p>We then explored how GNNs translate raw network data into mathematical vectors via Feature Learning. Unlike traditional models, GNNs utilize Message Passing to learn node representations (embeddings) that encapsulate both a node’s own attributes and its structural context within the graph.</p><ul><li><strong>Neighborhood Aggregation:</strong> Each node “polls” its neighbors, combining their features using functions like Mean, Sum, or Max Pooling to capture local patterns.</li><li><strong>Recursive Learning:</strong> By stacking layers, a node can “reach” further neighbors, allowing it to learn global information from across the network.</li><li><strong>Representation Space:</strong> The goal is to project nodes into a low-dimensional space where nodes with similar roles or connections are clustered together, enabling tasks like link prediction or community detection.</li></ul><p>We then learned about the different types of Graph Neural Networks (GNNs), focusing on how each architecture is designed to handle specific graph characteristics and learning tasks. This included spectral-based and spatial-based GNNs, as well as popular models such as Graph Convolutional Networks (GCNs), Graph Attention Networks (GATs), and GraphSAGE. Each approach differs in how it aggregates and propagates information across nodes, allowing GNNs to effectively model complex relationships in both homogeneous and heterogeneous graphs.</p><p>Overall, the session highlighted how Graph Neural Networks provide a powerful and flexible framework for learning from interconnected data, enabling more accurate and context-aware intelligence in domains where relationships are as important as the data itself.</p><p><strong>Conclusion</strong></p><p>Our journey through this Winter School revealed that the future of AI lies in the balance between complexity and accountability. From the biological networks of the brain to the relational depth of GNNs, we have explored how intelligence is structured. Yet, as models ingest more of our “data exhaust,” technical safeguards like XAI and Differential Privacy move from being optional features to foundational requirements.</p><p>Ultimately, a “Smart” future is only viable if it is transparent. By bridging the gap between high-performance algorithms and human-centric values, we ensure that the next generation of AI is not just a powerful tool, but a trustworthy partner in our social and clinical world.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9d12c8cb2a29" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>