Projects

Problem

Diagnosing Kubernetes incidents requires cross-referencing pod events, logs, node pressure, resource quotas, and deployment state — repetitive work that takes time and deep context every time something breaks.

Approach

Built a ReAct agent loop with LangGraph on the Claude API, backed by 15+ diagnostic tools across seven modules — pods, logs, events, nodes, workloads, mutations, and OpenShift-native resources. Mutating actions pause for explicit human approval before any write executes. Multi-cluster routing via JWT claims, async SSE streaming from FastAPI, and a real-time Next.js dashboard with an AI chat console, cluster topology view, and node heatmap. Ships with Helm, Argo CD GitOps, and a Prometheus/Grafana observability stack.

Outcome

A production-grade agent that reasons through CrashLoopBackOff, pending pods, quota exhaustion, and rollout failures — and acts on them safely, with a full approval audit trail and built-in observability.

Problem

Handing an AI agent real cloud power is dangerous by default: raw credentials, shell access, and irreversible operations don't mix with a system that can be wrong.

Approach

The agent never holds cloud credentials or shell access — it only issues MCP tool calls to purpose-built servers wrapping a narrow, auditable set of operations, with every call written to an audit log. Terraform modules for Azure networking, AKS, and IAM plus Ansible; an MCP infra server wrapping terraform plan/apply/destroy; and a LangGraph + FastAPI agent core that requires human confirmation before any apply or destroy.

Outcome

A safety architecture for agentic DevOps: narrow tool surfaces, full audit trails, and a human gate in front of every irreversible action.

Problem

Engineering context lives scattered across tickets, PRs, docs, and meetings — so planning is manual, knowledge walks out the door, and delivery risk is discovered late.

Approach

Built with a collaborator: LangGraph + RAG agents over the team's own artifacts, an event backbone on Kafka, vector retrieval with Qdrant, and a multi-provider LLM layer. Python 3.12, FastAPI, PostgreSQL 16, MongoDB, Redis, Next.js 14, Socket.IO, Keycloak, MCP, Docker, and GitHub Actions.

Outcome

Sprints planned from real signals instead of guesswork — PRDs and meetings become tickets, and the knowledge base answers from the team's own history.

Problem

Debugging CI failures is slow and context-heavy. Engineers manually cross-reference logs, run history, and code changes just to tell a real regression from a flake.

Approach

TypeScript in strict mode with no any, bundled with esbuild for the Node 24 action runtime. Deterministic flakiness heuristics (same-SHA pass/fail, transition counts) run before the Anthropic SDK is called with structured outputs — no free-text parsing — plus prompt caching and prompt-injection sanitization, because CI logs are untrusted input. PR comments are idempotent: later failures update the existing comment.

Outcome

CI triage in seconds — the PR comment surfaces cause, suggested fix, and a flakiness confidence score before an engineer opens the logs.

Problem

A transport operation needed a reliable, secure ERP it could roll out progressively — without re-architecting at every stage from a developer's laptop to the cloud.

Approach

Multi-environment path from Docker Compose on Rocky Linux through OpenShift CRC to Azure Red Hat OpenShift using Kustomize overlays. Nginx over TLS to Odoo to PostgreSQL 15 with HAProxy load balancing, MFA enforcement, delivery automated with Argo CD (GitOps), and observability with Grafana and Loki. Stateless services with externally managed configuration and secrets.

Outcome

A repeatable deployment pipeline that promotes the same artifacts across environments, with production-grade reliability and security built in from day one.

Problem

Most document chatbots hallucinate beyond their sources and give no way to verify where an answer came from.

Approach

LangGraph agent on Claude Haiku with tool-call routing, checkpointed memory, and token streaming. Qdrant with hybrid retrieval — dense plus BM25 fused with RRF — and a cross-encoder rerank stage. Local fastembed embeddings, so there's no second API key and it's reproducible offline. FastAPI backend, Next.js 15 frontend with HTTP and SSE streaming.

Outcome

Grounded answers with clickable citations, honest refusals, and scope limits documented rather than hidden. Built as a technical assessment.

Problem

DevOps is hard to learn from reading — it needs real clusters, real pipelines, and the chance to break and fix things safely.

Approach

Container-first design with a Go backend, Angular frontend, and PostgreSQL persistence. Scenario-based labs (including Kubernetes) give each learner an isolated sandbox modelled on real CI/CD and incident-troubleshooting workflows.

Outcome

Guided learning paths tied to executable environments, so skill-building happens by doing rather than watching.

Problem

External systems emit events on their own schedule; downstream consumers need them validated, processed, and observable without coupling to the source.

Approach

Webhook ingestion with optional HMAC-SHA256 authentication, a Redis queue, and async workflow processing in Java and Spring Boot. Shipped with a public backend, Swagger UI, health checks, a Prometheus metrics endpoint, and a dashboard.

Outcome

A deployed, monitorable integration layer — built as an interview project and shipped like a real one.

Problem

Students needed answers grounded in real institutional documents and a structured read on their own CVs — not generic chatbot advice.

Approach

A FastAPI service called internally from a Spring Boot backend. Google Gemini 2.0 Flash for inference, BAAI/bge-m3 local multilingual embeddings via sentence-transformers, and retrieval over Qdrant and pgvector.

Outcome

Grounded institutional Q&A plus CV analysis and per-student learning roadmaps, running as a service inside the campus platform.

Problem

The annual career forum ran on manual, error-prone coordination across company registration, booth allocation, and contracts.

Approach

Built a platform that automates company registration, booth booking, and contract workflows end-to-end, alongside the club's internal tooling.

Outcome

Streamlined the forum for 20+ participating companies and replaced manual coordination with an automated workflow.

Problem

Content teams had no objective, repeatable way to judge whether an article was optimized for search.

Approach

Built independently as sole developer with Spring Boot, Next.js, and MongoDB. Article-scoring workflows combine web scraping, NLP, and OpenAI API signals across 500+ articles into a practical optimization score.

Outcome

Turned fuzzy content-quality signals into concrete scores teams could act on.