Huzaifa Shafique

National Textile University, Faisalabad

Bachelor of Science · Artificial Intelligence

August 1, 2022 – June 30, 2026

National Textile University, Faisalabad

Teaching Assistant Computer Organization & Assembly Language

February 1, 2025 – June 1, 2025

Faisalabad, Punjab, Pakistan

Sound Emotion Recognition Dual-Approach Audio Pipeline

January 1, 2026 – January 1, 2026

Built two independent SER pipelines and benchmarked them head-to-head: Approach 1 (40-dim MFCC + CNN-LSTM) reached 75% accuracy; Approach 2 (3-channel Mel spectrogram stacking Mel, Δ, and ΔΔ into a physics-aware 94×128×3 image + Bi-LSTM + custom Attention) reached 80% accuracy — a 5-point gain by encoding temporal dynamics as image channels rather than sequence features. Approach 2 uses rectangular CNN kernels (5×3 time-focus, 3×5 frequency-focus) and a custom Attention layer that suppresses silence and padding frames, focusing the model on voiced segments. Both pipelines use Mixed Float16 precision, light audio augmentation (Gaussian noise, ±1 semitone pitch shift, 0.95-1.05× time stretch), and WarmUp + Cosine Decay scheduling.

Tell2Design Generative AI for Architectural Floor Plans

January 1, 2026 – January 1, 2026

Developed a generative pipeline that transforms natural language room descriptions into structured 2D floor plan layouts using Graph Attention Networks (GAT) trained on the 3D-FRONT indoor scene dataset (18,000+ professionally designed rooms). Modeled spatial dependencies between rooms as a graph — nodes represent rooms, edges encode adjacency constraints — enabling the model to generate layouts that respect real-world architectural relationships (e.g. kitchen adjacent to dining, bedroom separate from living areas). Evaluated layout coherence with IoU-based room overlap metrics; demonstrated that GAT-based spatial reasoning outperforms sequence-to-sequence baselines on constraint satisfaction for multi-room plans.

Facial Emotion Recognition Blendshape MLP Pipeline

January 1, 2026 – January 1, 2026

Explored three architectures progressively before arriving at the production model: (1) FaceNet backbone fine-tuning with 3-phase curriculum training; (2) CNN/ViT baselines; (3) a compact 165-dim temporal blendshape MLP — chosen for its combination of accuracy, pose invariance, and mobile deployability. Refactored from a 1404-dim raw landmark feature space to a 52-dim FACS-aligned blendshape representation, achieving a 3x model size reduction with higher pose invariance; expanded to 165-dim by appending temporal derivatives and 4 geometric ratios (EAR, MAR, brow-to-eye, mouth pull). Architecture: GLU blocks + Residual MLP + Multi-Head Self-Attention with Supervised Contrastive Loss (T=0.07) + Focal Loss (y=2); trained on AffectNet + RAF-DB with AdamW + Cosine Decay Restarts. Achieved 78% validation accuracy on AffectNet and 68% zero-shot transfer on RAF-DB (held out entirely during training), demonstrating generalization beyond dataset-specific texture cues. Exported to TFLite INT8 with representative-dataset calibration; quantization metadata JSON enables direct drop-in to Android MediaPipe pipelines.

Virtual HR Multimodal Mock Interview Analysis System

January 1, 2025 – January 1, 2026

Built a real-time Android application that conducts AI-powered mock interviews: generates job-specific questions via LLM, records candidate responses, and produces a structured multimodal feedback report covering verbal fluency, facial emotion, and posture – all running entirely on-device with no server dependency. Integrated Whisper (speech-to-text) + BERT (fluency and coherence scoring) for verbal analysis; MediaPipe Pose for posture detection; and the custom 165-dim blendshape MLP (see FER project) for real-time facial emotion recognition – three parallel inference streams in a single APK. Deployed full inference stack via TFLite INT8 quantization, achieving sub-100 ms on-device latency across all three modules on mid-range Android hardware. (AI/ML backend and model integration by author; Flutter frontend by collaborator.) Delivered a complete, installable APK – not a prototype – with per-module structured feedback output designed for repeat use by job seekers.

Edge AI Livestock Tracking & Classification System

January 1, 2025 – January 1, 2025

Developed an edge-based animal tracking and classification system using ESP32-S3, ESP32-CAM, and TensorFlow Lite; dual IR-sensor direction detection triggers real-time image capture and on-device inference with a quantized MobileNet V1 model classifying livestock (Cow, Goat, Hen) with no cloud dependency. Logged classification events and running counts to InfluxDB via Wi-Fi; added OLED status display and a Streamlit interface for validating the TFLite model on uploaded images during development.

YouTube Video Question Answering RAG Pipeline

January 1, 2024 – January 1, 2024

Built a Streamlit application that answers natural language questions over any YouTube video — extracts the transcript, chunks and indexes it with FAISS + all-MiniLM-L6-v2 embeddings, then generates context-grounded answers via Meta LLAMA 3 (LangChain + Together AI). Supports multi-language transcripts and configurable retrieval size k; includes chunk-level source inspection so users can verify which segments informed each answer — addressing the hallucination transparency problem common in naive RAG deployments. Handles edge cases gracefully: missing captions, API rate limits, and videos with auto-generated vs. human transcripts are all caught and surfaced to the user with actionable error messages.