In The Living Algorithm, Shafaq Bajwa argues that inclusive artificial intelligence requires immersion in real classroom contexts, not just technical optimization. Drawing on her transition from software engineering to special education, she highlights the human factors that define technology adoption and learning outcomes. Bajwa critiques the EdTech industry’s focus on abstraction and scalability, calling for empathy, flexibility, and co-creation with neurodiverse learners. Her perspective blends data science, education, and ethics into a call for AI that reflects human complexity rather than oversimplified models.
The tech industry’s push for personalized education through AI often overlooks a fundamental gap: the chasm between clean algorithms and the messy, dynamic reality of a classroom. For truly inclusive AI to emerge, product development must move beyond technical specifications and immerse itself in the lived experiences of diverse learners, especially those with special educational needs.
Shafaq Bajwa, a data science postgraduate with deep roots in software engineering, embodies this intersection. After a career that included roles as a Junior Analyst Programmer and Computer Science Lecturer, she now serves as a Learning Support Assistant at Moorcroft School in London. This transition from coding banking applications to directly supporting students with complex needs provides her with a crucial perspective on the limitations of technology built in isolation.
To create technology that serves all students, EdTech companies must look beyond traditional technical qualifications. Bajwa advocates for immersion programs that embed developers in real-world educational settings, arguing that direct exposure is essential for building empathetic and effective solutions.
She states, “To build truly inclusive AI, EdTech companies should go beyond hiring for technical skills and create immersion programs that bring real-world, lived experiences into product development. This could include embedding team members in under-resourced schools, rotating engineers into frontline roles like support or teaching, or hiring neurodivergent learners and students from marginalized backgrounds as paid co-designers.” This approach aligns with research on hybrid methodologies that combine ethnographic insights with machine learning to inform context-aware technology.
Bajwa adds, “These programs help teams understand diverse learning contexts and prevent biased or one-size-fits-all solutions. These approaches ensure products reflect the needs of all learners—not just the most represented ones.” This hands-on experience is critical for developing systems that can interpret complex human interactions, a challenge addressed by models like GraphMFT, which uses graph networks for multimodal emotion recognition.
For ventures targeting neurodiverse learners, traditional metrics like user acquisition and scalability fail to capture the full scope of their impact. Bajwa argues for a redefinition of ‘return on investment’ (ROI) that prioritizes meaningful social and educational outcomes over short-term financial gains.
“Techno-Stars would refine return on investment by shifting the focus from purely financial gains to meaningful educational and social outcomes. For neurodiverse learners, ROI would be measured in improved confidence, increased engagement, and progress in learning goals—outcomes that may not be easily quantified but have a lifelong impact,” she explains, referencing her startup project. This perspective is vital in a sector where some private equity firms have focused heavily on realized value through sales and recapitalizations.
By helping learners access education tailored to their needs, she says, the value extends beyond the individual. “This broader definition of ROI emphasizes long-term human potential and inclusion, rather than short-term profit alone.” This aligns with the goals of impact investing, where venture capital funds like Sun SEA Capital are backing startups in the EdTech and HealthTech sectors to drive sustainable growth.
Classrooms generate a unique form of ‘relational data’ derived from the dynamic interactions between a student, a teacher, and an educational tool. Unlike structured data, this information is fluid and context-dependent, posing a significant challenge to the architecture of conventional machine learning models.
Bajwa notes, “This is a more complex relational form of data that we would have to build machine learning models to address, as opposed to structured data, where the interaction between student, tool, and teacher is more static and less influenced by context, emotions, and teaching styles.” Capturing these nuanced interactions requires sophisticated approaches, such as models that can dynamically learn from conversational structures, a key feature of the GASMER framework for emotion recognition.
The goal, she suggests, “To build models where inputs [are not] isolated from each other, but where it is important to build more adaptive context-aware systems that can respond to real-time classroom interactions and learning needs more effectively.” This ambition is reflected in advanced research exploring how to capture the dynamics of emotional changes over time using methods like the Dynamic Graph Neural Ordinary Differential Equation Network (DGODE).
Drawing on her experience as a computer science lecturer, Bajwa emphasizes the need for data science education to confront real-world ambiguity. She proposes a project that would challenge students to build assistive technology for non-verbal students with special educational needs.
“The challenge would be to create a prototype that can take limited inputs, such as eye movement, touch patterns, or simple gestures, and produce outputs, such as words or emotions,” she explains. Such a task would force students to work with messy, incomplete data, a problem that often requires a deep understanding of the user’s context, as highlighted in studies on data annotation accuracy.
“This project would teach students that real-world problems often involve messy, incomplete data, ethical considerations, user-centered design, and understanding human behavior and accessibility, not just algorithms,” Bajwa explains. This underscores that success in educational technology is not solely about code but about achieving real impact in complex human environments.
A significant chasm exists between how technology is designed and how it is adopted in complex human environments. Bajwa’s transition from developing structured banking applications to facilitating learning in a special needs classroom has illuminated this ‘implementation gap’ firsthand.
“The biggest gap I have seen between technology and implementation is that technologies are designed for ideal users and ideal workflows, but real human environments are messy and unpredictable,” she observes. This reality is a central challenge in fields like federated learning, where models must perform robustly across diverse and imperfect real-world settings.
She adds that adoption hinges on factors of comfort, trust, and accessibility, noting, “People do not always use tools as designed, but adapt, resist, or find ways around them. It is not about building a perfect product, but about designing with empathy, ongoing support, and flexibility.” This requires building systems that can handle the unpredictability of human behavior and protect user privacy, a core concern in frameworks that manage dynamic differential privacy.
Terms like ‘personalization’ and ‘adaptive learning’ are ubiquitous in the tech industry, but their application often falls short of their promise. From her frontline perspective, Bajwa sees these concepts frequently reduced to simple algorithmic adjustments that miss the deeper needs of a student.
“From my frontline experience, terms like ‘personalization’ and ‘adaptive learning’ often get reduced to simple algorithm tweaks or content recommendations, missing the deeper reality,” she says. This superficial approach tends to treat students as data points rather than whole people with emotions, context, and changing needs throughout the day, a challenge that multimodal emotion recognition frameworks aim to address.
True personalization, Bajwa argues, requires understanding that a learner’s state is constantly in flux. She concludes, “Human support, empathy, and flexibility are just as important as technology in meeting those moment-to-moment needs,” an insight that aligns with research into using multimodal data fusion to gain better semantic context for human activity recognition.
A fundamental principle in software engineering is abstraction—hiding complexity behind a simple interface to create scalable products. However, daily immersion in the unabstracted reality of student needs challenges this core tenet, revealing its potential to oversimplify the human experience.
“The raw reality of student needs at Moorcroft proves that abstraction, though necessary to build scalable tech, can sometimes oversimplify the human experience,” Bajwa reflects. “Seeing the full complexity of learners’ emotions, behaviors, and support needs makes it clear that hiding behind simple interfaces risks missing critical context.” This issue is what data fusion techniques in multimodal AI attempt to solve, such as those in the Sync-TVA framework.
“This challenges the industry to design products that balance abstraction with flexibility and personalization and human judgment rather than one-size-fits-all solutions,” she adds. It’s a call to create systems that can handle nuance, a key focus in developing AI that leverages rich, multimodal information.
Reflecting on her career, which began as a Junior Analyst Programmer, Bajwa offers an insight that has reshaped her trajectory. The nature of ‘systems’ extends far beyond code and processes to encompass the people and contexts they operate within.
“If I could go back in time and talk to my younger self, I would say that systems are not solely about code or processes, but about the people who are using them and the contexts in which they are operating,” she states. This human-centric view is increasingly important in designing privacy-preserving AI, where frameworks must account for user data sensitivity across different modalities, as seen in Heterogeneous Privacy Federated Learning (HPFL).
Understanding that “technology is part of a larger human and organizational ecosystem” is key. This perspective, she notes, “To do more meaningful work from day one by focusing on solutions that meet user needs, not just technical specifications that sometimes neglect issues like communication overhead, which is a focus of Compressed Private Aggregation methods.”
Bajwa’s journey suggests that the future of inclusive technology depends on co-creation within the classroom. This approach requires blending technical expertise with direct human insight, aiming to build systems that are both effective and contextually aware.
