Selected Research-to-Production Work
Each project demonstrates the full research-to-production loop: identifying the AI problem that matters, designing the architecture and evaluation system, building or guiding the implementation, measuring whether it works, and turning the resulting knowledge into reusable systems, patents, papers, and technical decision frameworks.
AI Innovation: Patents, Publications & Research Foundations (2016-now)
Research-to-impact evidence spanning Machine Learning, Deep Learning, GenAI, NLP, Speech AI, multilingual AI systems, semantic retrieval, benchmark design, evaluation science, terminology intelligence, big data management, domain adaptation, Agentic AI, KGs, and production-grade AI quality improvement for enterprise systems.
Portfolio:
US patents and research publications covering semantic domain assignment, terminology disambiguation, abbreviation resolution, specialised-vocabulary ASR evaluation, benchmark design, data-centric evaluation, and adaptive STT fidelity methods for reducing hallucinations and skipped content.
Foundations:
Academic work in word sense induction, semantic clustering, neural data-to-text generation, semantic summarisation, retrieval QA, and deep learning–based task-oriented dialogue management.
Enterprise Speech AI: Zero → Production (2024-2026, SAP SE)
How do you build reliable enterprise Speech AI when no prior product, benchmark, or evaluation infrastructure exists?
Defined and built SAP's multilingual Speech AI capability from initial ambiguity to enterprise production across ASR, TTS, speech translation, voice workflows, domain adaptation, voice cloning, robotics-oriented speech optimisation, and speech data product architecture.
Role:
Lead AI & Applied Data Scientist / Sole AI Expert
Scope:
ASR · TTS · Speech Translation · Voice Workflows · Evaluation · Production
Key Contributions:
Designed the applied research direction, evaluation architecture, model-selection logic, and quality-improvement methodology for enterprise Speech AI. Built terminology-aware adaptation paths, integrated automated metrics with human evaluation and robustness checks, and developed decision logic linking offline quality signals to deployment readiness and improvement priorities.
Challenging Reality:
No prior product or benchmark infrastructure; multilingual scope across enterprise domains; sensitive audio data; legal, licensing, data-protection, procurement, and supplier constraints.
Evidence:
Production launch in April 2026; up to 20% WER improvement on selected SAP-domain evaluation settings; approximately 10% MIM improvement on selected terminology/content-completeness scenarios; pending patent on STT fidelity, hallucination, and skipped-content reduction; co-authored LREC-COLING 2026 paper on specialised vocabulary evaluation for ASR.
Methods & Technologies:
ASR · TTS · Speech Translation · Voice Cloning · Fine-tuning · STT/TTS/E2E Evaluation · Terminology-aware Adaptation · OVKWS-style Terminology Injection · LLM-based Post-editing · G2P/Text Engineering · Noise Robustness · Human + Automated Evaluation · Domain-specific Quality Logic
TTS Evaluation & Pronunciation Quality Framework (2026, SAP SE)
How do you evaluate enterprise TTS when the critical failures are code-switching, unknown Latin-script terminology, abbreviations, pronunciation accuracy, and naturalness — not just generic audio quality?
Designed and implemented an automated TTS evaluation framework for enterprise speech generation, reducing dependency on large-scale human evaluation while making terminology-specific pronunciation quality measurable under real product constraints.
Role:
Lead AI & Applied Data Scientist / Evaluation Architect
Key Contributions:
Created the architecture, methodology, benchmark logic, and implementation for terminology-aware TTS evaluation. Focused the framework on code-switching of Latin-script terminology, unknown words, abbreviations, mixed-script inputs, pronunciation correctness, pause behaviour, intelligibility, and naturalness. Combined recogniser-based term checks, pronunciation-oriented proxy metrics, reference-free quality estimation, targeted expert validation, and ablation-based comparison of preprocessing and pronunciation-control strategies.
Methods & Technologies:
TTS Evaluation · Code-switching Evaluation · Latin-script Terminology · Unknown-word Evaluation · Abbreviation Handling · ASR/Recognizer-based Term Recognition · Term Reading Accuracy · Kana/Phoneme/Mora Edit Distance · Pause-duration Analysis · Generation-failure Tracking · Reference-free Quality Estimation · UTMOSv2 · SpeechBERTScore / Discrete Speech Metrics · GOP / CTC-GOP Research Extension · Human Expert Validation · Ablation-based Evaluation
Terminology-Aware Enterprise AI Architecture: RAG, KGs & Agentic AI (2019-now, SAP SE)
How do you design enterprise GenAI systems when terminology, multilingual knowledge, governance, retrieval quality, and workflow reliability matter more than generic chatbot capability?
Shaped terminology-aware AI architecture and feasibility directions across RAG, knowledge graphs, agentic workflows, query reduction, and LLM-based/hybrid enterprise systems, grounded in SAP's large-scale multilingual terminology and language-data infrastructure.
Role:
Lead AI & Applied Data Scientist / AI Architect / Enterprise AI Advisor / AI Ambassador
Scope:
Enterprise Terminology Intelligence · RAG · Knowledge Graphs · Agentic AI · Query Reduction · LLM/Hybrid Systems · Evaluation Workflows
Key Contributions:
Connected large-scale terminology intelligence with practical enterprise AI architecture. Analysed and structured SAP terminology and multilingual language assets, including 12M+ multilingual term pairs, identifying ambiguity, abbreviation, metadata, bilingualism, synonym, and semantic-domain issues that limited downstream AI reuse. Co-authored patents on semantic domain assignment, terminology disambiguation, and abbreviation resolution, turning enterprise language-data problems into reusable AI/IP assets. Extended this foundation into GenAI and hybrid-AI feasibility work: advised on RAG, knowledge-graph-enhanced architectures, agentic workflow patterns, SAPTerm-based use cases, query reduction, and terminology-aware support-answer retrieval. Led or shaped AI-side feasibility, architecture, evaluation logic, reliability criteria, data requirements, and continuation boundaries for initiatives including the GlobalizeMe Query Reduction work and follow-on internal Agentic AI exploration.
Evidence:
Screened and shaped a portfolio of 100+ internal AI automation ideas; contributed as invited AI expert in leadership-level AI discussions and workshops; provided AI advisory across Speech AI, LLMs, RAG, KGs, and agentic systems. Patent-backed terminology work includes semantic domain assignment, terminology disambiguation, and abbreviation detection/mapping. AI architecture work covered query reduction, retrieval/generation trade-offs, tool-use boundaries, workflow orchestration, hallucination risk, data readiness, and evaluation design for LLM-based enterprise workflows.
Methods & Technologies:
Enterprise Terminology Intelligence · RAG Architecture · Knowledge Graphs · Agentic Workflow Design · Query Reduction · LLM-based Systems · Hybrid AI Architecture · Semantic Domain Assignment · Terminology Disambiguation · Abbreviation Resolution · Retrieval Evaluation · Hallucination Risk Analysis · Workflow Evaluation Logic · AI Feasibility Assessment · Enterprise AI Advisory
OVKWS for ASR with Custom Vocabulary & Synthetic Benchmarks (2025, SAP SE)
Can customer-specific terminology be recognised in enterprise ASR without large domain-specific fine-tuning corpora?
Originated and led applied research with UNISINOS on customer-specific terminology recognition for enterprise ASR.
Role:
Lead AI & Applied Data Scientist / Applied Research Lead / Research Direction Owner
Scope:
ASR · OVKWS · Custom Vocabulary · Synthetic Data · Benchmark Design · Multilingual Evaluation
Key Contributions:
Originated the research direction and turned the challenge into a structured applied research initiative. Defined research hypotheses, methodological boundaries, success criteria, and evaluation logic. Co-designed the OVKWS-based architecture, synthetic-audio strategy, and specialised-vocabulary benchmark. Coordinated the research as Lead AI Scientist and co-authored the resulting publication.
Evidence:
Led research direction for a private enterprise OVKWS-based approach to custom vocabulary recognition in ASR, combining open-vocabulary keyword spotting, synthetic audio, acoustic matching, and hard-negative strategies. The system demonstrated substantial quality improvements on selected company-domain specialised-vocabulary benchmarks; detailed implementation and system results remain internal. Co-authored LREC-COLING 2026 paper: A Dataset for Evaluating ASR on Specialised Vocabulary. Introduced a linguistically curated bilingual benchmark for English and Portuguese, comprising 13,846 utterances / 18.7 hours, distributed across synthetic and literature-derived subsets, with OOV rates reaching up to 100%. Proposed a diagnostic evaluation framework separating Biased Word Error Rate, targeting domain-specific jargon, from Unbiased Word Error Rate, focused on general vocabulary. Results: Up to +25% MRR improvement on selected English company-domain specialised-vocabulary benchmarks and up to +133% MRR improvement on selected Portuguese company-domain benchmarks.
Methods & Technologies:
ASR · Open-vocabulary Keyword Spotting · Custom Vocabulary Recognition · TTS-generated Synthetic Audio · Cross-attention for Acoustic Matching · Hard-negative Sampling · Low-data Adaptation · Specialised-vocabulary Benchmarks
Multilingual Speech Data Product (2025, SAP SE)
How do you turn fragmented private enterprise audio sources into reusable, governed AI assets for speech systems?
Architected the data-centric foundation for SAP's multilingual Speech AI capability: a governed speech data product transforming raw, fragmented, private enterprise audio into reusable AI-ready assets for Speech AI workflows under enterprise constraints such as licensing, consent, governance, and traceable reuse.
Role:
Lead AI & Data Scientist / Speech Data Product Architect
Scope:
Speech Data · Data-centric AI · Metadata Architecture · Governance · Multilingual Scale
Key Contributions:
Defined the architecture, data model, metadata logic, and reuse strategy for an internal multilingual speech data product. Designed entity relationships across raw audio, segmented audio, transcripts, translations, speakers, sessions, terminology annotations, and processing artefacts. Structured a 20+ stage pipeline covering ingestion, segmentation, alignment, preprocessing, transformation, metadata engineering, validation, deduplication, licensing checks, and dataset assembly.
Evidence:
Architecture for a 20+ stage processing pipeline; metadata and governance model with 100+ dimensions covering modality, interaction type, language, domain, consent, and processing metadata; designed for approximately 40 languages and multiple downstream speech AI workflows. Designed storage architecture using AWS S3 and SAP DataSphere.
Methods & Technologies:
Speech Data Architecture · Data-centric AI · Metadata Modelling · Pipeline Architecture · Amazon S3 · SAP DataSphere
ASR Evaluation & Quality Governance Framework (2024-2025, SAP SE)
What makes an ASR evaluation framework representative enough to support real deployment decisions?
Built a reusable evaluation and benchmarking architecture for enterprise ASR, connecting general transcription quality, domain-specific robustness, terminology recognition, dataset representativeness, and deployment-readiness logic across multilingual enterprise scenarios.
Role:
Lead AI & Applied Data Scientist / Evaluation Architect
Scope:
ASR · Enterprise Speech Recognition · Benchmarking · Domain-specific Evaluation · Terminology Evaluation · Quality Governance
Key Contributions:
Designed and implemented the ASR evaluation methodology, benchmark structure, quality dimensions, metadata logic, automated evaluation pipelines, human-evaluation integration, robustness checks, and decision logic connecting offline metrics to production readiness. Structured evaluation assets across approximately 90 datasets and 8+ languages, with 100+ metadata and quality dimensions covering language, domain, audio type, interaction type, noise, accents, terminology density, modality, content completeness, and product-relevant scenarios. Contributed to up to 20% WER improvement in selected enterprise ASR scenarios and to patent-pending stabilisation concepts for hallucination and skipped-content reduction.
Methods & Technologies:
ASR Evaluation · Benchmark Representativeness · Multi-dimensional Dataset Characterisation · Domain-specific ASR Evaluation · Terminology-specific Evaluation · Automated Metrics · Error Analysis · Metric Correlation · Quality Governance
Semantic Retrieval & MLTR Evaluation Framework (2019-2024, SAP SE)
How do you evaluate multilingual semantic retrieval quality at enterprise scale, when relevance, translation quality, similarity, and downstream usefulness all matter?
Designed and implemented the evaluation backbone for SAP's Multilingual Translation Repository (MLTR), a large-scale enterprise multilingual retrieval platform within SAP Translation Hub, operating across millions of verified translations, 40+ target languages, 2,000+ language combinations, and thousands of SAP customers.
Role:
Applied AI & Data Scientist
Scope:
Semantic Retrieval · MLTR · NMT · Translation Quality · Multilingual Evaluation · Enterprise NLP
Key Contributions:
Owned the AI-facing design of retrieval relevance, evaluation architecture, quality measurement, scoring logic, experimentation, and improvement direction. Co-shaped the transition from legacy keyword-based retrieval to semantic retrieval, designed multidimensional evaluation across retrieval coverage, similarity, edit-distance indicators, translation-quality metrics, embedding-based similarity, and behaviour-oriented test scenarios simulating real translator usage.
Evidence:
Built scalable benchmarking workflows across language pairs, query types, search configurations, proposal scenarios, text lengths, domains, and representative datasets — a high-dimensional evaluation space far beyond a fixed set of reports. Created repeatable measurement infrastructure for systematic experimentation, quality diagnosis, and decision support across a multilingual enterprise retrieval platform used in large-scale SAP localisation workflows.
Platform Impact:
Supported the AI and evaluation foundation behind a platform used in large-scale customer rollouts, including contexts involving 80k+ text elements and 25k+ custom objects. SAP customer case studies for related platform usage reported up to 93% reduction in manual translation effort, 99%+ proposal accuracy, 87% faster translation timelines, and 1,250 hours of developer effort saved.
Methods & Technologies:
Semantic Retrieval Evaluation · Multilingual Benchmarking · Retrieval Relevance · Translation Quality Metrics · BLEU · COMET · ChaacTER · Embedding Similarity · Edit-distance Metrics · Quality Scoring · Confidence Scoring · Representative Test Sets · Benchmark Automation
Japanese Enterprise TTS Optimisation without Extensive Fine-Tuning (2024, SAP SE)
How can Japanese enterprise TTS handle unknown Latin-script terms, abbreviations, and code-switching naturally when extensive fine-tuning is not feasible?
Designed, co-implemented and validated a pipeline-based approach for improving Japanese TTS pronunciation and naturalness in terminology-heavy SAP contexts, focusing on unknown words, Latin-script product terminology, abbreviations, code-switching, Katakana conversion, text engineering, and pronunciation-oriented preprocessing. Mentored the student in Japan conducting hands-on experiments and native-speaker validation.
Role:
Lead AI & Applied Data Scientist / Applied AI Research Lead / Mentor
Scope:
Japanese TTS · Code-switching · Unknown Words · Latin-script Terminology · Pronunciation Quality · Non-fine-tuning Optimisation
Key Contributions:
Defined the research direction, solution architecture, and evaluation methodology for terminology-aware Japanese TTS optimisation. Designed the pipeline-based approach across lexicon preprocessing, Katakana conversion, phonetic normalisation, abbreviation handling, pronunciation rules, and targeted text-engineering strategies. Guided implementation and experimentation, created detailed evaluation logic.
Evidence:
Substantial improvement in targeted SAP terminology pronunciation and naturalness on selected SAP-domain test scenarios for Japanese. Early evaluation showed up to 64% WER improvement in relevant settings. Co-authored patent under review: "Enhancing Pronunciation Accuracy and Naturalness in Japanese Text-to-Speech Systems without Extensive Fine-tuning."
Methods & Technologies:
Japanese TTS · Code-switching Handling · Unknown-word Processing · Latin-script Terminology · Katakana Conversion · Text Engineering · Lexicon Preprocessing · Phonetic Normalisation · Abbreviation Handling
Deep AI Foundations: ML, Deep Learning, NLP & Conversational AI (2014-2018, Heidelberg University, Empolis Information Management)
How do foundational NLP and deep learning systems built before the GenAI era inform today's LLM, RAG, and agentic AI work?
Built and evaluated pre-GenAI NLP and Deep Learning systems across semantic representation, unsupervised clustering, neural generation, extractive summarisation, retrieval-based QA, task-oriented dialogue management, information extraction, and domain-specific search — providing the technical foundation for modern LLM, RAG, and hybrid AI architectures.
Role:
AI Scientist / Computational Linguist / Researcher / AI Developer
Scope:
Machine Learning · Deep Learning · NLP · Conversational AI · Semantic Representation · Pre-GenAI Research
Key Contributions:
Designed and implemented complete experimental pipelines from preprocessing to modelling, evaluation, and error analysis. Built unsupervised word sense induction and semantic clustering systems using sense2vec-style representations, vector-mixture embeddings, and MeanShift clustering. Developed neural data-to-text generation with encoder–decoder architectures and LSTM decoding. Extended LexRank-style extractive summarisation with semantic-similarity features. Built task-oriented conversational AI combining retrieval-based and neural approaches for technical-support QA.
Evidence:
CHERTOY word sense induction system based on the SemEval-2013 WSI task; neural data-to-text generation using MS COCO, V-COCO, and COCO-a; LexRank-based semantic summarisation; and Heidelberg University thesis on deep learning dialogue management for task-oriented conversational agents. Industry NLP foundation: At Empolis Information Management, contributed to applied NLP and smart-data projects including information extraction, text mining, sentiment analysis, linguistic analysis, and an NLP-driven scientific paper search engine in the pharmaceutical domain. Why it matters now: These foundations support stronger judgment in modern AI systems: when to use retrieval instead of generation, how to evaluate semantic similarity, how to debug representation failures, how to design hybrid architectures, and when simpler NLP/ML methods may outperform heavier LLM pipelines.
Methods & Technologies:
Machine Learning · Deep Learning · NLP · Word Sense Induction · Semantic Clustering · sense2vec Representations · MeanShift Clustering · Neural Data-to-Text · Encoder–Decoder Architectures · LSTM Networks · LexRank Summarisation · Semantic Similarity · Information Extraction · Text Mining · Scientific Search · Sentiment Analysis · Retrieval-based QA · Task-oriented Dialogue Management