RAG chatbot over an author’s writing
Answers questions from Ramazan Demir’s published writing, and refuses when his sources don’t cover it. I built the retrieval side, and every prompt change has to pass 40 hard benchmark questions first.
Principal Software / Platform Engineer · Python & Applied AI
I build Python services and the platforms they run on. On the AI side I’ve shipped a RAG chatbot that answers in Turkish and built context routing for a coding agent. Right now I’m building a coding agent with its own independent grader, to get better at measuring whether any of this works.
I BUILD IT,Answers questions from Ramazan Demir’s published writing, and refuses when his sources don’t cover it. I built the retrieval side, and every prompt change has to pass 40 hard benchmark questions first.
A typed classifier picks which context a coding agent needs before it opens any files. It never grants permission to change anything.
A local coding agent fixes bugs; a separate container grades each patch with hidden tests. The first baseline passed 15 of 15, so the tasks were too easy. Harder ones are next.
QLoRA on Llama 3.1 8B, trained on a 16 GB MacBook and graded by my own checker. Neither model invented a figure; the fine-tune hit the house format 50 of 50 times, against 0 for the prompt alone.
A client’s answer robot needed to speak. I built the data pipeline from raw recordings to transcribed, tested training splits, picked speakers by pitch, and trained a multi-speaker VITS model on rented GPUs.
Principal Software Engineer, via Mobiquity
I own two FastAPI services other teams depend on: an orchestrator and the layer that stores its data. Mostly that means API contracts, performance, and environments that build the same way on a laptop and in the pipeline.
Explore the services ↗Principal Engineer from Jul 2023, Senior+ Platform Engineer before that
My first four years at Mobiquity were on this project, from day one. I rebuilt the AWS platform in Terraform, put logs from 10 production accounts in one place, load-tested before setting the autoscaling, and built the RAG chatbot on datakeeper.nl.
Centralized logging ↗Explore the platform ↗The longer version, if you want it. Pick a node or a file.
I build Python services and the platforms they run on, with ownership of contracts, performance, and delivery.
Principal engineer on Brit Insurance, through Mobiquity. I own two FastAPI services: the orchestrator, and the layer that stores the data.
A coding-agent lab with a separate grader. It’s where I practise evals, fine-tuning, retrieval testing, MCP, and running models locally.
Four years on Datakeeper, a Rabobank identity wallet. Before that, DevOps for Vatan Bilgisayar, a Turkish electronics store: Kubernetes, caching, pipelines.
Write-ups with the actual code: RAG context limits, typed routing for a coding agent, a QLoRA fine-tune, voice training, and a few platform pieces.
OPEN THE NOTEBOOKIf you’re hiring for software, platform, or applied AI engineering, LinkedIn is the quickest way to reach me.
Contact Mert on LinkedIn ↗Explore my GitHub ↗