Scalable differentiation, private inference, and structured language.
Parallel and Distributed Computer Systems research at POLITEHNICA Bucharest, applied AI engineering at the Research Institute, and production ML at Bitdefender. Four directions, spanning published results and ongoing work.
Published2025
Scalable High-Order Automatic Differentiation
Problem
High-order derivatives (Hessians, and beyond) are central to PDE-constrained deep learning, but computing them at scale is memory- and compute-prohibitive with standard automatic differentiation.
Approach
Building on the Stochastic Taylor Derivative Estimator (STDE) method (Shi et al., NeurIPS 2024), implemented a multi-GPU JAX version using Taylor-mode AD, randomized jet sampling, and pjit/pmap primitives for cross-GPU parallelism, then applied it to distributed, fault-aware settings.
Result
A 6.86× runtime speedup and 27% lower GPU memory footprint across 4×V100 GPUs, with near-linear strong scaling and precision to machine epsilon, validated on high-dimensional many-body Schrödinger PDEs. Also presented a parallel/distributed formulation at RoEduNet 2025.
Running ML inference on sensitive data (health, finance, government) typically requires exposing plaintext data to the model host, creating a privacy and compliance liability.
Approach
Studied homomorphic encryption schemes for ML inference, encrypting inputs so a model can compute on them without ever decrypting the underlying data.
Result
Published findings on the accuracy/latency trade-offs of homomorphic-encryption-based inference in the Annals of the Academy of Romanian Scientists (2025).
Turning unstructured or spoken language (meeting audio, documents, natural-language queries) into structured, queryable, trustworthy information is still brittle, especially under constrained (non-frontier) open-weight models.
Approach
At the Research Institute, built LLM-based knowledge-graph extraction pipelines (NER, relation extraction, structured prompting), RAG pipelines for domain-specific QA, and natural-language-to-structured-query systems. Separately, built a speaker-attributed meeting-transcription refinement pipeline combining diarization, ASR, and constrained open-weight LLMs.
Result
The meeting-transcription refinement work was published in Future Generation Computer Systems (Elsevier, 2026), evaluated on the AMI Meeting Corpus. The knowledge-graph and RAG systems are in active use in applied AI research systems at the Research Institute.
Reliable speaker diarization and voice-activity detection are prerequisites for almost everything downstream in speech systems, but production audio is noisy and diverse.
Approach
Benchmarked voice-activity-detection systems across 1M+ audio clips, built a custom CRDNN detection model, and later applied CNN-based speech-activity detection and Pyannote/Whisper-based diarization pipelines in production at Bitdefender.
Result
A CRDNN model surpassing SOTA VAD baselines by 1.4% (99.8% detection accuracy); a production diarization/STT pipeline that took manual QA effort from 120h/month to 5h/month processing 49k+ calls.