AI Research Engineer  ·  Founder, INFERENCE Lab  ·  Low-Resource NLP  ·  Speech Intelligence  ·  LLM Engineering

AI Research Engineer and founder of INFERENCE Lab, an applied AI research and engineering organization based in Multan, Pakistan.

My work is end-to-end: dataset construction, architecture design, model development, evaluation, and deployment as usable software. I work across low-resource NLP, speech intelligence, LLM systems, and production AI deployment — owning the full lifecycle from raw data through published research and live inference APIs.

All research is released publicly with reproducible pipelines, deployable inference code, and permanent DOIs.

Datasets, model weights, and reproducible artifacts are released alongside each paper.

RUEmoCorp: A Large-Scale Roman Urdu Corpus and Benchmark Suite for Emotion Classification Under Review — Language Resources and Evaluation (Springer) · Preprint: ResearchSquare (10.21203/rs.3.rs-9759243/v1) Construction of the largest Roman Urdu emotion recognition corpus to date. Introduces a cross-institute annotation validation framework with structured annotator roles, multi-round calibration, and inter-annotator agreement measurement. Benchmarks multiple model architectures. Dataset released on HuggingFace and Harvard Dataverse.

Continuous Vocal Load Monitoring in Professional Voice Users: Development and Occupational Validation of an Automated Assessment Tool Under Review — Journal of Voice A deployable occupational health monitoring system for professional voice users. Addresses the gap between laboratory vocal fatigue research and real-world monitoring tools. Validated against occupational use conditions. Built on the ECAPA-TDNN-VHE encoder.

Modeling Vocal Fatigue as Embedding-Space Deviation Using Contrastively Trained ECAPA-TDNNs Preprint · DOI: 10.5281/zenodo.18366305 · Under Review — Springer EURASIP Journal on Advances in Signal Processing Novel framing of vocal fatigue detection as deviation measurement in speaker embedding space. Contrastively trained ECAPA-TDNN-VHE achieves 2.5× improvement over the standard baseline. Multi-language, multi-microphone dataset collected independently across 90–100 speakers.

Data-Centric Roman Urdu NLP: High-Quality Dataset Curation, Privacy-Preserving Embeddings, and State-of-the-Art Model Benchmarking Preprint · DOI: 10.5281/zenodo.18080524 Comprehensive data-centric study of Roman Urdu NLP infrastructure gaps. Covers dataset curation methodology, privacy-preserving embedding strategies, and systematic model benchmarking. State-of-the-art sentiment classifier: 0.84 accuracy, 0.83 Macro-F1.

RUDaSA: Roman Urdu Dataset for Sentiment Analysis — A Large-Scale, Curated Corpus with Privacy-Preserving Embeddings and Competitive Benchmarking of Transformer Models DOI: 10.21203/rs.3.rs-9827763/v1 RUDaSA is a large-scale Roman Urdu sentiment analysis benchmark that provides privacy-preserving embeddings and evaluates state-of-the-art transformer models to advance NLP research for low-resource and code-mixed languages.

SecureCipher v2.0: 10-Dimensional Hyperchaotic Image Encryption with ECC Key Generation and Double HMAC Authentication Manuscript in preparation Benchmarked against 14 encryption algorithms including DNA+Chaos combinations and AES-based systems. SecureCipher v2.0 scored 100/100 across all evaluation criteria: entropy 7.9993, chi² 244, NPCR 99.61%, with double HMAC authentication and confirmed lossless reconstruction. Next highest-scoring system scored 70/100.

Forecast-Based Decision Support System for Mango Malformation Disease Preprint · DOI: 10.5281/zenodo.16090477 · Ahmad, M.K. & Mangana, A.A. Time-series forecasting pipeline and smart agriculture decision system demonstrating 50–60% yield improvement through data-driven interventions.

QueryVault — LLM Observability & Hallucination Detection End-to-end LLM observability system with real-time hallucination detection, structured logging, and a developer dashboard tracking hallucination rate, P50/P95 latency, and error events.

Auralis — Production REST API for Vocal Health Containerized REST API accepting raw audio and returning normalized vocal fatigue scores via ECAPA-TDNN-VHE. Deployed on HuggingFace Spaces.

VigilantEye — Real-Time Violence Detection Real-time violence detection on live camera feed using a fine-tuned deep learning pipeline with instant alert triggering and frame-level logging. Optimized for low-latency inference.

Confidence & Posture Analysis — Behavioral Intelligence Multimodal desktop application combining computer vision and NLP for real-time speaker confidence and posture analysis via webcam. Generates structured post-session reports.

INFERENCE Lab is an applied AI research and engineering organization I founded in Multan, Pakistan. The lab conducts original research in low-resource NLP and speech intelligence, builds custom AI systems for deployment, and runs structured engineering education programs for developers who need to move from syntax-level Python to production AI systems.

All lab research is released publicly on HuggingFace and Harvard Dataverse with permanent DOIs and reproducible pipelines.

Core

Speech & Audio

NLP & LLMs

LLM Engineering

Computer Vision

Deployment & Infrastructure

Databases

_{All published models, datasets, and libraries are available on this profile and at huggingface.co/Khubaib01}