Fine-Tune, Prompt, or RAG?
A practical decision frame from production work: when behavior must be learned vs when answers must track live, permissioned data.
Coming soon
FULL STACKAI ENGINEER
Production RAG · Agents · Python · TypeScript · Go
©2026
/Shipping since 2020
Hey!
I'm Noman — a full-stack software engineer in Finland who ships LLM systems customers run in production, not just demos.
At Aktual Oy I work embedded with client teams: production RAG on PostgreSQL and pgvector (hybrid search, reranking, permission-aware retrieval), agentic platforms with evals and cost monitoring on Azure, and full-stack product work — React dashboards, APIs, Temporal, and CI/CD behind it all.
The same foundation shows up everywhere else — Django and FastAPI services, Go microservices on Kafka, React Native apps, and the testing and review habits that keep releases boring in a good way. M.Sc. in Computer Engineering from Åbo Akademi University.
I fine-tune when prompts can't hold the behavior (LoRA / QLoRA); use RAG when answers depend on live, permissioned data; and measure what shipped — faster reports, lower token spend, and reliability you can put in a slide.
From idea to production. RAG assistants, agentic workflows, and full-stack services on Python, TypeScript, and Go — built for measurable quality, latency, and cost, with the same rigor as any backend system.
Full-stackengineeringfirst,thenAIthatearnsitsplaceinproduction.RAGpipelines,agents,andtheAPIs,datalayers,andevalsthatkeepthemfast,permission-safe,andworthrunningafterthepilotends.
Services
Production AI & RAG
- pgvector & hybrid search
- Agents & tool calling
- Evals & observability
Agentic Systems
- LangGraph / LangChain
- Temporal workflows
- Token & cost efficiency
Backend & APIs
- FastAPI & Django
- REST design
- Event-driven (Kafka)
Full-Stack Product
- Next.js & React
- React Native
- Dashboards & integrations
Cloud & DevOps
- Azure & AWS
- Docker & Kubernetes
- CI/CD pipelines
LLM Fine-Tuning
- LoRA & QLoRA
- Unsloth
- When to fine-tune vs RAG
Selected work
Featured projects
View all workThoughts
Notes on building
Short pieces on production AI, full-stack architecture, and the habits that keep systems maintainable.
Evals That Survive Production
LLM-as-judge, human-authored cases, and regression tests — measuring whether a change improved answers and at what cost.
Coming soon
The Role of Event-Driven Design in Microservices
Why message brokers like Kafka help teams create resilient, decoupled systems at scale — still the backbone under many AI pipelines.
Coming soon
Contact
Let’s work
together
Tell me what you’re building. I read every message and reply within a day or two.