Projects

A growing suite of AI agents that automate repetitive daily engineering and research tasks — turning hours of manual work into seconds of autonomous execution. Agents handle tasks such as automated code review & PR summarisation, literature monitoring (daily ArXiv digest filtered by interest), meeting preparation (agenda generation, pre-read summaries), data pipeline health checks, and report drafting from raw experiment logs. Built with multi-model orchestration using Claude Code and Gemini, observability via Langfuse, and security guardrails covering LLM security (prompt injection defence, output validation, audit logging).

LLMs range from a few GBs to hundreds of GBs — far exceeding the memory available on edge devices with shared DRAM (CPU, GPU, and accelerators contending). Working with Dr. Ismet Dagli to develop efficient memory management mechanisms that eliminate bottlenecks and enable LLMs / LVMs to run reliably on NVIDIA Jetson and similar edge hardware.

A voice AI product that empowers meeting secretaries with real-time transcription, speaker diarization, content summarization, and semantic search — built on LLMs and state-of-the-art audio processing research.

Full-stack face recognition solution: cloud-based registration, an extensive identity database, and on-edge inference. Co-developed with Oxford University advisers — achieving >99.4% accuracy, 30 FPS, and support for >10,000 identifiers. Deployed in hospitality and regulated gaming sectors.

In-house framework that automates quality review for an image stock platform across 20+ quality standards, processing up to 30,000 images/day with a 70% reduction in operational cost and 99.5% precision. Handles tiny object detection, domain diversity, and can substitute human reviewers in well-defined decision scenarios.

Adapted Stable Diffusion to generate commercial-grade stock photography via fine-tuning, LoRA, and responsible-AI constraints. Delivered a unified app supporting txt2img, img2img, inpaint, outpaint, upscale, and image variation — all within a single optimized serving pipeline.

Built an ML re-ranking layer on top of PIXTA's search engine, trained on billions of user interaction signals, resulting in a ~6% revenue increase (~$30,000/month). Deployed with full MLOps automation: data ingestion → training → evaluation → rollout.

Multi-label classification model for 25,000+ visual categories, serving both user-facing tag suggestions and internal dataset annotation. Evolved from ResNest (2020) → Meshed-Memory Transformer (2021) → pure Vision Transformer, with training on 5M+ samples.

Multi-task model built on a large face recognition backbone, predicting age and emotion while applying context-aware recognition to fuse facial and scene-level features — significantly improving label quality on a noisy, in-the-wild dataset.

Clustered hundreds of thousands of unlabeled face images to automatically assign consistent identity labels, accelerating dataset construction for downstream training. Overcame challenges including extreme class count, varied attributes, non-human faces, and label noise.

Built a proof-of-concept voice chatbot for Sorgenia, an innovative Italian energy company — the first voice bot in the energy sector in Italy. Developed at Pi School, Rome using Dialogflow and custom NLU pipelines.

Designed and built the core NLP platform for Panasonic smart devices — covering speaker identification, speech recognition, wake-word detection, and a multi-domain dialog system. Deployed in a pilot at a healthcare center for elderly residents in Kyoto, Japan.