Faiaz Amin Khan

I research and build systems that span the physical-digital divide. From designing patent-grade acoustic sensing mechanism at Samsung R&D to deploying AI workflows in resource contrained device, my work focuses on decoding human signals to create seamless, intelligent interactions.

🚀 Overview

I architected and prototyped an activity-aware recommendation system that fuses physiological sensing (PPG) with physical context (IMU) to deliver personalized health insights.

  • Status: Showcased at Samsung AI POC Exhibition (Internal R&D Showcase).
  • Role: Ideation Lead & Algorithm Engineer.

❓ Why?

Wearables today suffer from a “Data Rich, Insight Poor” paradox. They are excellent at logging 10,000 steps, but often fail to tell the user what to do with that data.

I realized that for a wearable to be a true health companion, it needs to shift from a passive “Quantified Self” tracker to an active “Guided Self” agent—one that anticipates user needs before they even ask.

📉 The Problem

Building a recommendation engine for a wrist-worn device presented three unique HCI challenges:

  1. The Cold Start: How to provide meaningful advice to a brand-new user with zero historical data?
  2. Context Blindness: A high heart rate means ‘exercise’ in the gym, but ‘stress’ in a board meeting. Simple sensors often fail to distinguish these states.
  3. Notification Fatigue: Generic “Time to move!” alerts are often ignored. We needed hyper-personalization to ensure engagement.

💡 The Approach

Due to Samsung Research IP regulations, specific algorithmic weights and proprietary logic are omitted.

I designed a Multi-Layer Data Fusion engine to determine user intent:

  • Physiological Layer: Aggregates heart rate variability and intensity metrics.
  • Physical Layer: Analyzes IMU (inertial measurement unit) data to classify micro-activities.
  • Situational Layer: Leverages device state to infer social context (ensuring data privacy by processing entirely on-device).

The Hybrid Inference Strategy: To solve the “Cold Start” problem, I implemented a multimodal learning model. The system initially bootstraps using heuristic clinical best practices, then seamlessly transitions to collaborative filtering as user interaction data accumulates, progressively tailoring the timing of insights to individual behavior.

🛠️ My Contributions

I owned the logic design and collaborated on the Proof of Concept (PoC) development:

  • Algorithm Design: Designed the ‘Activity Context’ logic that differentiates between ‘sedentary work’ (mental focus) and ‘sedentary relaxation’ (leisure), allowing for context-appropriate interventions.
  • Data Strategy: Curated a baseline dataset to power the initial heuristic models.
  • Edge Optimization: Engineered the inference engine to run efficiently on resource-constrained wearable APs (Application Processors), ensuring zero impact on battery life.

🏆 Impact

  • Recognition: Selected for showcase at the Samsung AI POC Exhibition, demonstrating the potential for next-gen health features.
  • Thought Leadership: Helped shape internal R&D thinking around “context-aware health intelligibility”—moving the product strategy from simple data logging to actionable intelligence.

🧠 Retrospective: Key Learnings

Navigating the complexity of this system taught me to synthesize diverse sensor streams into coherent user context.

  • Adaptive Systems: I learned that an effective recommendation engine must balance immediate utility (heuristics) with long-term personalization (learning).
  • Communication: Showcasing this work at the AI Exhibition honed my ability to translate complex sensor fusion concepts into compelling narratives for diverse stakeholders—a capability I view as essential for impactful research.

Updated: