Tutor.mk
Personalised AI learning platform
- Macedonian curriculum
- Local payments
- 320 + students
Open ↗
Featured projects
Evaluation highlight
LLM evaluation in VM-based TEEs - under 20 % performance cost on a 110 M-param BERT.
What the four panels show & how we produced them
My contribution to the study was a fully reproducible Bash + Python pipeline that (1) launches Kata-Containers under AMD SEV-SNP, (2) runs a parametrised train.py/infer.py for every ⟨batch size, epoch⟩ pair, and (3) ships raw logs for aggregation. The four graphs to the right are auto-generated via Matplotlib as the pipeline's final step.
- Top-left – TTFT (ms): start-up is priciest (~17 % on small batches) as the VM negotiates keys and encrypts 768 GB of RAM.
- Top-right – TPS: once warm, throughput sits ≈ 14 % below baseline — encryption bandwidth, not compute, is the bottleneck.
- Bottom-left – QPS: query-level throughput mirrors TPS; no latent spikes in single-tenant runs.
- Bottom-right – End-to-end latency: curves rise linearly with batch size, gap stays flat — overhead is stable.
Net takeaway: a 110 M-param BERT can train & serve entirely inside a VM-based TEE with a predictable 10-16 % tax.
CV & Tech Stack
Front-end & UI
- React · Next.js 14 (CSR/SSR/ISR, App Router)
- TypeScript ‑ strict typing, custom hooks, context
- Tailwind CSS, theming, accessibility, responsive design
- Live2D / Canvas graphics & image optimisation
Real-time media
- Browser voice calls that connect in <200 ms, powered by a custom WebRTC stack
- Adaptive noise suppression, volume levelling & live caption overlay for accessibility
- Auto-reconnect & cross-device hand-off keep conversations running 99.5 % of the time
Back-end & Databases
- Node.js / Express / Fastify · Python
- Next.js API routes & Firebase Functions
- PostgreSQL (Drizzle ORM) · Firestore · Redis
- Stripe webhooks, billing, rate-limiting, auth
AI / LLM tooling
- Hugging Face Transformers · PyTorch (MPS/GPU) · Fine-tuning
- Function / tool calling with JSON-schema definitions
- Prompt engineering · Retrieval & memory protocols
- Evaluation & benchmarking: scikit-learn metrics, latency & throughput
- Anthropic Claude & OpenAI Realtime/Assistants APIs
- Kata Containers (AMD SEV-SNP) secure model training & inference
- Streaming transcripts, TTS, vision & image tools
- Batch processing, usage metering & cost analytics
DevOps & Cloud
- AWS (EKS), GCP
- Docker · Kubernetes · Crossplane · Helm
- Terraform, Cloudformation, GitHub Actions / GitLab CI/CD, ArgoCD
- Vercel, Firebase, SSL/DNS, monitoring
Observability & Security
- Posthog · Prometheus · Grafana · Dynatrace · Datadog · ELK
- Hashicorp Vault · Kyverno · OPA/Gatekeeper
- Logging, alerting & incident response workflows
Leadership & Soft Skills
- Product-driven, UX-focused mindset
- Crisp technical writing & documentation
- Cross-functional collaboration & stakeholder communication
- Analytical problem-solving & decisive execution
- Languages: English (C2), German (B1), Macedonian (Native)
Timeline
- 2025 – present Co-founder / CTO, TwoPeas.ai
- 2025 – present Co-founder / CTO, Tutorist.ai
- 2025 – present Solo founder, Tutor.mk
- Apr 2024 – May 2025 DevOps Consultant (Contract), HanseMerkur Versicherungsgruppe
- Apr 2024 – May 2025 IT Consultant, Cloudsurf IT
- Aug 2019 – Jan 2023 Site Reliability Engineer, Joyn GmbH
- Jan 2019 – Aug 2019 Technical Operations Engineer, Joyn GmbH
Let's build something amzeballs
DMs open at @super_bavario or ping me on LinkedIn.