About
Hi, I’m Konstantinos, an AI Engineer at Deloitte. I design and build AI systems with a focus on automation, efficiency, and real-world impact turning complex problems into production-ready solutions that save time and reduce cost.
My work goes beyond generative AI applications. I have strong foundations in backend systems and machine learning, and I’m currently expanding into robotics and autonomous systems, bridging software intelligence with the physical world.
Experience
-
2025 — Present Build and deploy AI systems for large-scale enterprises within Deloitte, delivering production-ready solutions across multiple industries. Work closely with cross-functional teams including engineers, data scientists, product owners, and stakeholders to design, implement, and maintain reliable, scalable AI pipelines, while promoting best practices in system architecture, performance, and responsible AI.
-
Python
-
Fastapi
-
LangChain
-
LangGraph
-
Cloud
-
-
CanaryBench - On going
Built a red-team/blue-team platform for benchmarking prompt-injection defenses: one side hides a secret inside a system prompt, the other has a limited number of turns to extract it, with an LLM-based judge scoring each attempt. Designed the async backend with FastAPI, PostgreSQL, Redis, and an async task queue to handle matchmaking, authentication, rate limiting, and real-time scoring, with cost tracking across multiple LLM providers (OpenAI, Anthropic, Google). Used structured tool-use output for consistent, auditable evaluation results, and instrumented the platform with Prometheus and Grafana dashboards to turn match data into an attack-pattern dataset.
- Python
- FastAPI
- LangChain
- PostgreSQL
- Redis
- Docker
-
End-to-End MLOps Pipeline on AWS
Designed and built a production-style recycling image classifier (11 classes) on AWS SageMaker, fine-tuning a pretrained Vision Transformer and serving it through a real-time endpoint behind a REST API (API Gateway + Lambda). Implemented an automated SageMaker Pipeline (train → evaluate → conditional deploy) where an evaluation gate compares each candidate model against the one in production and triggers a zero-downtime blue/green deployment only if accuracy improves. Closed the loop with LLM-in-the-loop retraining: low-confidence production predictions are stored, automatically labeled by Claude through Amazon Bedrock, and, once enough valid samples accumulate, trigger a new training run. Tracked every training run with MLflow (metrics, parameters, and artifacts), monitored the system with CloudWatch custom metrics, and secured it with least-privilege IAM roles.
- Python
- PyTorch
- Hugging Face
- AWS SageMaker
- AWS Lambda
- Amazon Bedrock
- Amazon S3
- MLflow
- CloudWatch
-
Multimodal RAG Chatbot Platform
Built a full-stack platform where users upload PDFs and receive sourced answers: an async FastAPI backend with hybrid dense and keyword search via Qdrant over document chunks, plus PostgreSQL for structured data. Designed a semantic router that decides per query whether to search the user's documents or answer directly, and an Advanced Search mode that decomposes complex questions into parallel sub-queries and merges results. Added multimodal support through image captioning via Gemini so answers can draw on figures and diagrams inside PDFs, alongside JWT auth, per-user data isolation, and streaming chat responses.
- Python
- FastAPI
- LangGraph
- Qdrant
- OpenAI
- Gemini
- PostgreSQL
-
Robotic Manipulator
Developing an ongoing robotic manipulator project to deepen my understanding of robotics through hands-on system building. I’m working with ROS 2, expanding my C++ skills for robotic applications, integrating Arduino for low-level hardware control, and using MoveIt for motion planning and manipulation. The system is modeled with URDF and SRDF, continuously simulated in Gazebo, and visualized in RViz. I also integrated voice-based control, enabling motion triggers through Amazon Alexa using the Amazon SDK, allowing high-level voice commands to translate into planned robotic actions.
-
Python, C++
-
ROS2
-
Arduino
-
Gazeboo - Rviz
-
URDF - SRDF
-
-
VeriVista
I developed VeriVista as a solo project, building the entire backend and AI integration for a Chrome extension that provides real-time fact-checking of selected text on any webpage. I implemented LLM-powered analysis to deliver an instant trust score (0–100) with evidence from multiple reliable sources, including detailed claim-level breakdowns. The system supports multi-language verification (18 languages) and follows a privacy-first approach, processing only the selected text, making it a fully self-contained tool that helps users quickly and reliably assess content credibility.
-
Python
-
LangGraph
-
Fastapi
-
Tavily
-
-
AWS Bedrock Multi‑Agent System
Built a AWS Bedrock Multi‑Agent System as a solo backend project leveraging Amazon Bedrock’s multi‑agent collaboration capabilities to create an intelligent recommendation system for accommodation and restaurants. The system uses a supervisor agent and specialized collaborator agents that dynamically route natural language user queries to AWS Lambda functions, which process structured data from Amazon S3 and return results in real time. By orchestrating multiple agents with distinct responsibilities, the project demonstrates how multi‑agent AI can decompose complex tasks and deliver precise, context‑aware responses. The architecture combines AWS Bedrock Agents, AWS Lambda, Amazon S3, and boto3 to enable scalable, serverless AI workflows that showcase practical multi‑agent orchestration in production‑oriented environments.
-
Python
-
AWS Bedrock
-
AWS Lambda
-
Amazon S3
-