SANKARDEV S
Applied AI Engineer with 3+ years of experience building production NLP systems, transformer pipelines, and retrieval-augmented AI applications.
model = RoBERTa.from_pretrained(
"emotion-classifier"
)
result = model.predict(text)
▶ F1-Score: 95.08% · Classes: 28
🎓 Research Interests · Industrial PhD
Research Metrics
F1-Score
RoBERTa NLP · JDSIS 2026
Text Samples
GoEmotions + Hate Speech pipeline
Latency Reduced
ML training batching optimisation
CREMA-D Accuracy
Multimodal Wav2Vec2.0 model
Emotion Classes
EmoTract real-time classifier
Featured Project
Latest Work
KlarKI — Local-First EU AI Act & GDPR Compliance Auditor
Privacy-preserving bilingual AI compliance platform for German SMEs — zero external API calls, full on-device inference.
Macro F1
≥ 0.95
NER F1
≥ 0.95
RAG P@3
≥ 85%
Architected a 7-container local-first bilingual AI compliance platform for EU AI Act + GDPR auditing, designed for German SMEs with zero external API calls and full on-device inference.
Built an end-to-end NLP pipeline covering document ingestion → actor classification → Annex III applicability gating → hybrid RAG retrieval → deterministic multi-agent LLM reasoning → scored PDF/JSON compliance reports.
Fine-tuned spaCy NER and deepset/gbert-base classifiers, deployed via ONNX + NVIDIA Triton, achieving 50–100× faster inference with reliable fallback through deterministic regex, NLI validation, and ML decision trees.
Improved compliance evidence retrieval by replacing basic RAG with hybrid retrieval using BM25 + ChromaDB + Reciprocal Rank Fusion + cross-encoder re-ranking, significantly reducing hallucination risk and improving audit defensibility.
Built 200+ automated tests covering unit logic, golden dataset validation, RAG retrieval quality, hallucination detection, and adversarial regulatory cases.
Tech Stack
Technical Stack
Built for AI at Scale
The Journey
Professional milestones and academic achievements.
Research Assistant — NLP & Multimodal AI
- Built scalable NLP data pipeline processing 250,000+ text samples from GoEmotions and Hate Speech corpora.
- Fine-tuned RoBERTa (PyTorch, Hugging Face) for multi-class emotion recognition; achieved 95.08% F1-score, published in JDSIS (2026).
- Reduced model training latency by ~20% through batching optimisations across classical and transformer-based models.
- Developed SoundSense, a multimodal vocal anomaly detection system using Wav2Vec2.0; achieved 78.10% accuracy on CREMA-D, outperforming SVM/RF baselines by 30%.
Software Engineer
- Developed backend systems using Laravel and MySQL for multiple client applications across diverse industry sectors.
- Built responsive frontends using React, improving cross-device usability.
Software Associate
- Contributed to delivery and maintenance of 200+ bespoke software solutions across Python, PHP, and React stacks.
- Deployed predictive ML models reducing manual data processing and inference optimisation time by ~20%.
Education
MSc Artificial Intelligence (with Industrial Placement)
2023 – 2025University of East London, London, UK
BSc Computer Applications
2019 – 2022Kerala University, India
Selected Projects
Things I've Built
VoxChain — Real-Time LLM Voice Agent
Built a real-time voice AI system with bidirectional streaming (speech ↔ LLM ↔ audio). Deployed end-to-end agentic pipeline enabling low-latency conversational interaction.
SmartSafe — Multimodal Liveness Detection
Designed biometric attendance system combining facial recognition with audio liveness detection. Generated multimodal confidence score to mitigate replay and spoofing attacks.
SoundSense — Multimodal Emotion Detection
Cross-modal system detecting emotional inconsistencies between speech text and vocal tone. Achieved 78.10% accuracy, outperforming SVM/RF baselines by 30%.
Latest Publication
"NLP Framework to Safeguard Youngsters Online Using Advanced Transformer-Based Models"
Journal of Data Science and Intelligent Systems (JDSIS), Q1-ranked (Scopus indexed, SJR), 2026
Let's Connect
I'm currently open to new opportunities in Applied AI, NLP, and RAG Systems engineering. Whether you have a question or just want to say hi, I'll try my best to get back to you!
© 2026 Sankar Dev S. All rights reserved.