Applied AI engineering
LLM pipelines, agents and NLP systems built for real workloads, with evaluation, cost control and monitoring included from the start.
- LLM agents
- RAG
- Evaluation
- NLP
Applied AI studio
We turn language models into software people rely on.
BerkeML builds production ML systems for client teams anywhere in the world, and ships its own AI-native products, end to end.
CompanyBerkeML · sole proprietorship
Based inIstanbul, Türkiye
ClientsWorldwide · remote-first
Contactberke@berkeml.com
01 — What we do
Two halves of one studio.
We sell engineering to teams who need AI in production, and we build products of our own. Each side makes the other better: client work keeps us honest about reliability, and our products keep us fast.
Document intelligence
Extraction, classification and reconciliation for document-heavy industries. Messy scanned PDFs go in, and structured, verified data comes out.
- OCR
- Table extraction
- Classification
- Mortgage
AI-native products
We design, build and launch our own web and mobile apps, from idea to production. Each one has a model at its core.
- Mobile
- Web
- Voice AI
- Generative
02 — Client work
AI for the mortgage industry.
BerkeML is a contracted engineering partner to a company that automates mortgage and title workflows with AI.
Ongoing engagement · Remote
Turning mortgage paperwork into structured, verified data.
Mortgage files are hundreds of pages of scans, tables and disclosures. We build the ML behind workflows that read those files, understand them and act on them, so the lender's team only handles the exceptions.
- Table & field extractionPull structured values out of tables in scanned PDF documents.
- Document classificationMulti-class models that sort every page of a loan file.
- Label qualityEnsemble-based label cleaning to make training data more trustworthy.
- Agentic workflowsLLM agents that carry out processing, closing and post-closing steps.
03 — Products
Things we build for ourselves.
Our own products, designed, built and launched end to end.
Live · WebHealth
Promil Tracker
Track drinks and see an estimated blood alcohol level, calculated with the Widmark formula. No login is needed. Optional Google sign-in adds history and statistics across devices.
promiltracker.comLive · WebSocial
InfluencersBid
A competitive leaderboard where creators bid their way to the #1 spot. It's a playful experiment in attention, ranking and real-time bidding.
influencersbid.lol
04 — How we build
Your problem first. Then the model.
We don't start from a favourite stack. We start by listening, then build the system that gives you the highest accuracy and the best return for your budget.
01 · Listen
Understand the need
We learn your workflow, your data and what a correct answer is worth to your business before we write any code.
02 · Right-size
Fit the budget
The simplest approach that works wins. We choose models and architecture by accuracy per dollar, not by hype.
03 · Measure
Prove the accuracy
Every system ships with an evaluation set and clear metrics, so you can see how well it performs, not just hear about it.
04 · Optimise
Keep it profitable
After launch we cut inference cost and latency, and keep improving accuracy as real data comes in.
Our promise
The right system is accurate enough to trust, cheap enough to run, and pays for itself.
05 — Founder
An ML researcher who prefers shipping.
Berke Dilekoğlu founded BerkeML after years of building NLP and LLM systems for companies around the world. His background runs from deep-learning research on biological sequences to production document AI, and from search ranking at Huawei to agentic mortgage workflows today.
- NowBerkeML — FounderClient AI engineering for companies worldwide, plus the studio's own AI-native products.
- 2023 →Areal.ai — Senior NLP EngineerDocument extraction, classification and LLM workflows for mortgage automation.
- 2023Ginoa.io — ML EngineerDocument-grounded chatbot pipeline with automated evaluation and cost/performance tuning.
- 2021 – 2023Huawei — NLP Research EngineerApp tagging, query expansion and keyword extraction for AppGallery search.
- 2017 – 2018Ford Motor Company Turkey & Durham University — Visiting ResearcherComputer vision: 360° image stitching, object detection, thermal re-identification.
- EducationSabancı University — MSc & BSc Computer Science, BSc MechatronicsHigh Honor Scholarship. Thesis: SUMOnet, deep sequential prediction of SUMOylation sites.
- Open sourceLangChain contributorBatched document updates for ChromaDB (8× faster), plus vector-store similarity docs.
Bioinformatics · 2022
Discovering misannotated lncRNAs using deep learning training dynamics.
bioRxiv · 2023SUMOnet: Deep sequential prediction of SUMOylation sites.
IEEE ASYU · 2021The impact of ensemble learning in sentiment analysis under domain shift.
arXiv · 2021ML with HE: Privacy-preserving machine learning inferences for genome studies.
06 — Contact
Let's build
berke@berkeml.com
Let's build
something useful.
berke@berkeml.com
Hiring for AI engineering, document automation or an AI-native product build? Send a short note about the problem. Replies usually go out within one business day.