01Agentic systems2026 — Present
Giles.app
Making complex agents work together.
AI Engineer at Giles
Infrastructure for multi-agent research: asynchronous inference, coordinated tools, and resilient knowledge pipelines.
- Google ADK
- MCP
- Vertex AI
- Qdrant
- Cloud Tasks
- MongoDB ACID
Engineering details +
- Architected asynchronous LLM Queue middleware that decouples Google ADK agents from Vertex AI (Gemini 2.5/3.5) through Google Cloud Tasks, with a dead-letter-queue sweeper.
- Built hierarchical multi-agent trees with turn-scoped contexts and declarative MCP tool servers covering European HTA bodies: IQWiG, G-BA, HAS, and NICE.
- Engineered a cluster-global permit scheduler in Qdrant and resilient ETL pipelines.
United Kingdom / Remote
02Computer vision2025 — 2026
PointRef Inc.
From pixels to a decision on court.
AI Pickleball Officiant
Real-time pickleball officiating on native iOS, connecting camera geometry, ball trajectories, and on-device inference.
- 3D PnP
- Core ML
- Homography
- Swift iOS
- YOLO
Engineering details +
- Formulated monocular 2D-to-3D solvePnPRansac camera calibration and planar homography for metric 20 × 44 ft court registration.
- Designed parabolic bounce detection and paddle-contact rejection, then converted models to Core ML for real-time edge inference.
- Advised Mosalam Ebrahimi, formerly of Meta Reality Labs, on the computer-vision system.
Sports technology · Advisory
03Production AI2024 — 2025
Accuwiz Tech · PadStyler.com
Reimagining a room. Preserving its structure.
Virtual Staging Platform
A generative vision engine for virtual staging, refurnishing, and remodeling, with depth-aware control over room boundaries.
- SegFormer
- MiDaS Depth
- YOLOv11
- FastAPI
- Redis
- AWS S3
Engineering details +
- Delivered the core generative vision engine across virtual staging, wall and floor refurnishing, remodels, and other renovation modalities.
- Combined fine-tuned SegFormer, YOLOv11, and MiDaS depth with Boolean mask algebra to lock structural room boundaries.
- Productionized the system on AWS with FastAPI, Redis session management, and S3.
Generative vision
04Multi-agent RAG2024 — 2026
InstaComply · SelfCompl.ai
Turning compliance into a working system.
AI Architecture at SelfCompl.ai
A modular multi-agent retrieval system for enterprise audits, with human review and traceable compliance workflows.
- Multi-Agent
- LangChain
- Pinecone
- AWS Textract
- PostgreSQL
Engineering details +
- Architected SelfCompl.ai for GDPR, ISO-27001, and SOC2 audits with real-time compliance gateways and compliance-as-code.
- Built human-in-the-loop audit trails and feedback, and multi-LLM orchestration across Claude, GPT-4o, and Gemini.
- Authored more than five governance research papers alongside the engineering work.
Legal technology
05Medical researchFeb 2022 — 2023
Wiseyak Inc.
Making predictions more interpretable.
Medical Diagnosis & Video TransCNN
Medical-image classification, explainable diagnostic views, and a hybrid CNN–Transformer architecture for video.
- Grad-CAM
- CLIP
- TransCNN
- PyTorch
- Meta SAM
Engineering details +
- Joined as a Machine Learning Researcher in February 2022. Engineered chest X-ray report generation across ten diseases using CLIP embeddings and a multi-label classifier.
- Integrated Grad-CAM explainability into the radiologist cockpit and combined CLIP with GPT-2 for report generation on imbalanced data.
- Designed TransCNN for medical-video denoising and next-frame prediction, and built a plant-disease domain-adversarial neural network using Meta SAM.
Medical AI & Research
06Healthcare GraphRAG2025 — 2026
Accuwiz Tech
Connecting knowledge to clinical questions.
Signal Healthcare Ecosystem & Spatial CV
Graph-based retrieval for healthcare intelligence, stateless document understanding, and spatial vision for design tools.
- Neo4j GraphRAG
- Model Context Protocol
- Vision LLM
- Spatial Homography
Engineering details +
- Architected SignalDX using Neo4j GraphRAG, with an autonomous MedicareAgent over live MCP. SignalMD covered more than 100 biomarkers.
- Built SignalDoc, a stateless Vision LLM that classifies 22 document types with fully in-memory, zero-disk multipart buffering.
- Built Lightology perspective homography and a Cabinet Design computer-vision engine that unwarps kitchen photographs into 2D orthogonal blueprints.
SignalDX.ai · Lightology · Cabinet Design
