Dmytro – Python, FastAPI, React

HIPAA-compliant medical imaging data pipeline for a healthcare client of Keylabs. Cloud ingestion, encryption and de-identification of DICOM studies, with client-specific tag profiles and a full audit trail back to the rule that produced each de-identified record.

• Owned design and delivery of a fully automated ETL pipeline for cloud ingestion and anonymization of medical imaging data, cutting overall processing turnaround by 20%.
• Implemented de-identification with client-specific tag profiles, replacing ad-hoc rules that assigned different identifiers to the same patient across datasets.
• Automated extraction, encryption and anonymization of sensitive medical records end-to-end, raising pipeline throughput by 30%.
• Made pseudonymization deterministic and replayable — the same source identifier always maps to the same pseudonym — with a per-tag transformation log, so any de-identified study could be traced back to the rule that produced it.
• Ran the pipeline's AWS footprint inside a closed perimeter — VPC-isolated subnets, security groups and least-privilege IAM roles with no long-lived credentials — so protected health data never crossed the client's network boundary.
• Ran the engagement end-to-end: requirements, delivery and direct technical communication with the client's compliance and engineering leads.

Speech, translation, OCR and retrieval features inside Keylabs' AI data-annotation platform — a provider-agnostic inference layer with self-hosted models (Whisper, YOLO) as the default and managed APIs as fallback, plus an MCP server exposing platform capabilities to LLM clients.

• Built a provider-agnostic inference layer for speech, translation and OCR — self-hosted models as the default, managed APIs as fallback — so backends could be swapped without product-side changes.
• Wrote the product's MCP server, exposing platform capabilities to LLM clients.
• Built RAG pipelines over internal documentation to automate domain-specific lookups, removing recurring manual research steps from routine workflows.
• Fine-tuned YOLO on in-domain imagery to power model-assisted pre-labels, improving mAP@50 on a held-out set.
• Deployed the platform's AI services and MCP server on AWS — containerised on ECS with SQS-backed job handling.
• Automated internal workflows with n8n pipelines, Telegram bots and Python/Bash/PowerShell tooling.

Monitoring service that installs into a customer's Kubernetes cluster and tracks ML model performance across cluster nodes, with a dashboard covering model health, performance and release versioning. Also covered LLM output evaluation with quality guardrails and structured annotation pipelines.

• Partnered directly with project stakeholders to translate high-level business objectives into detailed technical specifications for core microservices.
• Developed backend services for LLM output evaluation, integrating quality guardrails and structured annotation pipelines to enable automated model testing across AI-powered features.
• Built interactive data visualisation components in React, using GraphQL for efficient data fetching and managing complex application state to deliver real-time monitoring insight on the dashboard.
• Extended dashboard APIs with monitoring logic in Google Cloud Functions, adding metrics such as drift detection and latency percentiles, and event logging.
• Owned delivery: packaged the service as cloud-agnostic Helm charts installable into customer clusters on any managed Kubernetes, validated on both GKE and AKS, shipped through ArgoCD GitOps onto Terraform-provisioned private GKE.
• Set up deployment to staging and integration environments, streamlined releases through CI/CD pipelines, and covered core business flows with E2E and integration tests using Playwright.

Platform for managing the lead generation and conversion process, built to streamline sales team workflows and turn prospects into closed business — customisable drag-and-drop dashboard, internal admin portal, role-based access and an AI assistant prototype.

• Built a customisable dashboard with drag-and-drop widgets, owning the full path from React UI through widget state persistence to the supporting API.
• Developed an internal admin portal for the operations team covering internal analytics and data management, and implemented the API business logic behind it.
• Extended the Firebase authentication and authorisation layer with role-based access control, protecting both pages and endpoints.
• Prototyped an AI assistant on the OpenAI API, building the accompanying UI and WebSocket layer for live responses.
• Set up the deployment infrastructure on Google Cloud Platform and fully automated releases through CI/CD pipelines.

Learning platform enabling organisations to create, manage and share interactive courses, quizzes and assessments — multi-tenant, subscription-based, running on AWS.

• Led backend development with Django and PostgreSQL, architecting scalable data models and optimised queries to handle high-concurrency platform traffic.
• Built a configurable quiz and assessment engine supporting multiple question types, automated scoring and detailed per-question feedback.
• Integrated Stripe for subscription payments and Auth0 for secure multi-tenant authentication with role-based access control.
• Implemented fine-grained permission management to ensure security and compliance across multi-tenant environments.
• Deployed and ran the platform on AWS — containerised Django on ECS with RDS PostgreSQL, and course media served from S3 behind CloudFront.