What we can build — shown, not claimed.

Problem

Training, comparing and deploying models by hand is slow and hard to reproduce. Teams need a repeatable path from training run to a scalable inference endpoint.

Approach

A Kubeflow pipeline trains several models in parallel with different learning rates, selects the most accurate one and deploys it automatically as an auto-scaling KServe inference service, with artifacts stored in MinIO.

Problem

Spinning up a complete team environment on Azure — networking, an AKS cluster, a private PostgreSQL database, Key Vault and access control — usually involves several tools and pipelines.

Approach

A single Kubernetes resource (a Crossplane claim) requests a full environment. A Composite Resource Definition and Composition encapsulate the infrastructure logic, and Crossplane reconciles every Azure resource from the Kubernetes control plane.

Problem

Applications need credentials, but storing secrets in code or plain manifests is risky, and many organisations use more than one secret manager.

Approach

One consistent pattern, three providers: each secret store is provisioned with Terraform, and the External Secrets Operator synchronises secrets from HashiCorp Vault, AWS Secrets Manager and Azure Key Vault into Kubernetes.

Problem

Queue-driven workloads often sit idle or fall behind because CPU and memory are poor signals for how much work is waiting.

Approach

Producer and consumer services for AWS SQS and Kafka, plus a CPU-intensive workload, deployed with KEDA ScaledObjects that scale on queue length, consumer lag or CPU load.

Problem

Streams of security log events are too noisy for fixed rules alone, but sending everything to an analyst does not scale.

Approach

A LangGraph supervisor lets an LLM decide which tools to call for each event — heuristic scoring, a SQL lookup of recent logins, IP reputation, web search — and produce a plain-English verdict, alerting Slack when risk is high.

Problem

Keyword search misses support tickets that describe the same issue in different words.

Approach

Ticket text is embedded with Sentence Transformers, stored in a Qdrant vector database and queried through a FastAPI service that returns the most similar past tickets.

Problem

Finding a specific moment in long videos means scrubbing manually through them.

Approach

Spoken content and visual scenes are indexed with VideoDB and queried through LlamaIndex retrievers, returning and streaming only the relevant segments.

Problem

Several data pipelines with different dependencies need to share one orchestrator without being packaged together.

Approach

A Docker Compose deployment in which the Dagster webserver and daemon load each pipeline from its own user-code container over gRPC, with PostgreSQL for run, schedule and event storage.

Problem

Turning separate ML scripts into a tracked, repeatable pipeline.

Approach

Dataset artifact, data processing and training steps chained into a ClearML pipeline built from tasks.