Selected work

Projects

Systems running in production, followed by research, academic and independent work: event-based vision, generative AI, robot autonomy, sensor hardware and LLM agent pipelines.

Systems in production

Designed, built and deployed

Agentic document inspector & orchestrator

In production

Sales & engineering documentation · Multi-agent orchestration

  • Agentic document inspection and orchestration system that identifies contract-critical requirements and subtle specification changes across large engineering and commercial documentation.
  • Automated cross-document consistency and dependency management, propagating changes from a single source document across all related project documentation.
  • AI-assisted generation, validation and continuous updating of engineering specifications, financial documentation, tender documents and technical schematics.
Multi-Agent SystemsAgentic RAG Document IntelligenceHuman-in-the-Loop Guardrails

Real estate automation platform

In production

Multimodal computer vision · End-to-end workflow

  • End-to-end AI workflow generating portal-ready property listings from structured data, adapting content to buyer profiles and target markets.
  • Multimodal computer vision pipeline for room classification, alternate-view generation and virtual staging, including furniture removal/addition and style adaptation.
  • Integrated property pre-valuation based on location, floor area and room characteristics.
  • Web interface orchestrating the full workflow from a customized property form to one-click generation of listings and visual assets.
Diffusion ModelsVirtual Staging Image ClassificationContent Generation Web Orchestration

Workforce management platform

In production

Cross-platform · Android & iOS · Vision-based check-in

  • Cross-platform workforce management and time-tracking system with Android/iOS apps and facial-detection-enabled kiosk check-in.
  • Intelligent monitoring for shift inconsistencies, excessive working periods and missing rest intervals.
  • AI-driven report automation reproducing company-specific templates and business logic from internal workforce data.
  • Digital fire register using real-time attendance data to identify missing employees and support emergency workflows.
Facial DetectionMobile Apps Report AutomationCompliance Monitoring
MSc thesis · TU Dortmund

Event-based logistics model, dataset & NeRF 3D reconstruction

Technische Universität Dortmund · FLW Chair

  • Self-supervised learning model for 3D prediction of logistic objects from event-camera data.
  • Automated preprocessing, segmentation and labelling of event streams, using LLMs to assist the segmentation process.
  • Built a logistics dataset and NeRF-based 3D assets to improve training.
  • Open-source algorithms replacing physical position markers safely.
+6.3 mAP
Object detection over the baseline
90%
Annotation storage reduction
Thesis Repository Video
Animated event-camera output showing detected logistic objects
3D reconstruction of a logistics pallet Annotated logistics scene used for training

Computer vision & deep learning

Garment recoloring system

Shootify

  • Deep learning pipeline for automated colour correction.
  • Three neural networks: a dual-encoder recoloring U-Net plus two segmentation U-Nets with cross-attention and multi-task learning.
  • Cross-attention bottleneck for selective colour transfer; skip connections preserve structure; synthetic degradation pipeline.
  • Production CLI with training/inference tooling, mixed-precision training and composite losses (L1, ΔE, perceptual, GAN).

Repository

Multimodal generative AI framework

A+

Final project · University of Oxford

  • Unified framework integrating autoencoders, VAEs, GANs, diffusion models and CLIP.
  • Combined inpainting, conditional generation, style transfer, text-to-image synthesis and multimodal search in a single pipeline.
  • Outlined applications for creative industries, digital content generation and visual search.

Project summary

Multimodal visual intelligence platform

A+

Final project · University of Oxford

  • Unified detection (Faster R-CNN), classification (ViT + Swin), image translation (Pix2Pix, CycleGAN), video analysis and vision-language retrieval (CLIP).
  • Developed a shared embedding space for multimodal queries — text → image, image → image, video snippets.
  • Proposed use cases in biomedical analysis, creative industries and multimedia search.

Project summary

Robotics & algorithms

Guidance & navigation of a cleaning robot

Master's project · UPM

  • Navigation and cleaning algorithms in a simulated apartment using the Apolo simulator.
  • Localization via Extended Kalman Filter (EKF), path planning with Voronoi + A*, and reactive obstacle avoidance.
  • User interface for room selection plus systematic zig-zag coverage.

Report (PDF)

ROS2 node optimization in Aerostack2

CVAR · UPM

  • Aerostack2 firmware analysis across standard, composable and lifecycle nodes.
  • Recommended migrating mission nodes to composable form for materially lower resource usage.
ROS2Composable NodesLifecycle Nodes

Automated Tetris with AI

Master's project · UPM

  • C++ implementation of Tetris with AI decision-making.
  • Three training approaches: competition-based learning, gradient optimization and a genetic algorithm.
  • Reached a maximum score of 1.8M points with predictive strategies, well beyond typical human play.

Report (PDF)

Hardware, IoT & safety-critical

Wearable impact detection system

Patent

Bachelor's Thesis · UPM

  • Low-power portable system embedded in wearable hardware to detect impacts and monitor user status via capacitive fabric sensors.
  • Developed both the hardware prototype and system architecture, culminating in a patent filed and published by UPM.
Wearable impact detection prototype

View patent

IoT wearables with TENGs

IMDEA Materials

  • Novel approach to fire detection and alert systems using new materials.
  • Smart wearables and a full TENG-based intelligent vest.
  • Proof of concept for IoT devices built from novel materials.
TENGLoRaWi-Fi Low-Power Sensors

Publication

Airbus SMART MRTT

Airbus Defence and Space

  • Verification and certification of DAL-A requirements (internal processing).
  • Advanced training in RTOS, FPGA and microelectronics.
Airbus MRTT air-refuelling system

AI agents & automation

Independent projects

Outreach agent

Industrial PhD & hiring research

  • End-to-end outreach pipeline using AI agents and LLMs to match companies against user-defined filters (sector, location, visa, research fit).
  • Automated corporate email discovery and personalized message generation with intent-aware prompts.
  • Multi-step reasoning: company profiling → contact extraction → message planning → LLM-crafted outreach with tone control.
  • Retries, guardrails, rate limits and logging for safe, consistent delivery.

Job intelligence pipeline

LinkedIn + LLM agents

  • Scraping and enrichment workflow building a company and role dossier (tech stack, research lines, visa notes).
  • Customized LLM producing a relevance index per opening — skills match, seniority, research fit, location constraints.
  • Concise per-role report (pros, cons, risks, next steps) with automated email digests and direct apply links.
  • Scheduling, de-duplication and alerting for continuous monitoring.

Business automation frameworks

Reusable frameworks

  • Reusable automation frameworks integrating LLMs with multi-step agents for business operations.
  • Focused on improving customer interaction, reducing manual workload and enabling AI-driven personalization.
  • Scalable design using orchestration tools, APIs and custom LLM pipelines adapted per domain.