class Soham:
"""From model training to production Kubernetes — I build the full stack."""
role = "Principal Software Engineer @ Red Hat"
focus = "AI/ML Engineering"
stack = ["Agents", "RAG", "Fine-tuning", "MCP", "MLOps", "K8s"]
ships = "AI systems that run at enterprise scale"ai_and_ml:
agents: LangGraph | LangChain | CrewAI | MCP Protocol
rag: Vectorless (reasoning-based) | Traditional (embeddings)
fine_tuning: QLoRA | PEFT | TRL | Synthetic data generation
llms: Claude | GPT | Gemini | Open-source (Llama, Mistral, Qwen)
evaluation: Langfuse | Custom benchmarks | Agent-specific metrics
security: OWASP Agentic AI | Guardrails | OPA
infrastructure:
orchestration: Kubernetes | OpenShift | Helm | Kustomize
iac: Terraform | Ansible
observability: OpenTelemetry | Prometheus | Grafana
ci_cd: GitHub Actions | Tekton | ArgoCD
languages: [Python, Go, Bash]
databases: [PostgreSQL, Snowflake, Redis, MongoDB, Elasticsearch, Neo4j, Qdrant]| Project | What it does | Why it matters |
|---|---|---|
| arcana | Kubernetes-native AI platform with CRDs for agent lifecycle, skills, guardrails, and FinOps | Unifies agent orchestration on K8s — the missing platform layer |
| toolglot | Universal IR for LLM tool calling — define once, export to any provider | Zero competitors. Solves tool-calling fragmentation across OpenAI/Anthropic/Gemini/Cohere |
| pageindex | Vectorless RAG using hierarchical indexing + LLM reasoning | No embeddings, no vector DB — just reasoning. A different approach to retrieval |
| owasp-agentic-scanner | Static analysis for OWASP Top 10 Agentic AI risks | pip install owasp-agentic-scanner — catch agent security issues before they ship |
| smithery | Fine-tune tool-calling models from your API definitions | QLoRA training on consumer GPUs. Your tools, your model, your accuracy |
| cognitive-memory | Biologically-inspired memory with intelligent forgetting for AI agents | 4-tier memory (working → episodic → semantic → procedural) with decay and consolidation |
@ Red Hat:
| Project | Stars | What it does |
|---|---|---|
| template-mcp-server | 54⭐ | Production-ready MCP server template with OAuth + OpenShift deployment |
| template-agent | 33⭐ | Enterprise LangGraph agent template with K8s security best practices |
From The Opinionated SRE:
From Medium:
- Why AI Agents Need Their Own Passports: The Engineering of Autonomous Identity
- SRE Odyssey: Navigating Chaos to Forge Collaborative Triumph
- Locked and Loaded: Fortifying AWS EC2 Security with an Ironclad Metadata Block
- Terraform: Aurora Postgres DB Upgrade
- Using ChatGPT and terraform
Building something in AI? Let's talk.