Applied AI Engineer | OCR, RAG, LLM Evaluation, Edge AI
Applied AI engineer for OCR, RAG, evaluation, and edge AI systems.
I help teams evaluate, build, and harden AI systems that have to work under real constraints: private data, noisy documents, inconsistent labels, and non-technical end users. My strongest work sits at the boundary between experimentation and production: OCR/LLM evaluation, retrieval quality, regression harnesses, and edge-model feasibility.
Recent delivery includes replacing Azure Document OCR with a local pipeline across 100+ invoice and vendor templates, adding schema-drift checks before deployment, and advising embedded-camera vision choices across Tiny/MobileViT, INT8, and accelerator options. At HKUST, I led research products that became usable systems, from NarrativeHive and Orchid to HK-GenSpeech and SeaSense.
Interactive Turing Machine
Serkan Kumyol
State
q0
Head
2
Action
S
Read ·, write S, move R
Output Register
S
Machine Type
Single-Tape Deterministic
Program Rules
13 transitions
What I Bring
Evaluation before deployment
I design benchmarks, regression checks, and failure criteria so teams can compare models on evidence instead of intuition.
Private-data document AI
I build OCR, extraction, and retrieval workflows for messy documents where security, controllability, and auditability matter.
Research translated into product
I take ideas from papers and labs and turn them into interfaces, demos, and internal tools that stakeholders can actually use.
Selected Work
See all projectsA hiring-manager cut of the work: recent delivery, systems that survived real user studies or evaluation gates, and projects that show how I operate under ambiguity, technical constraints, and stakeholder scrutiny.
Document Intelligence + Evaluation • 2026
Private OCR/LLM Evaluation for Invoice Processing
Benchmarked 100+ invoice and vendor templates to replace a cloud OCR dependency with a local privacy-preserving pipeline, then delivered the regression and deployment recommendations used for vendor selection.
Generative Narrative Systems • 2025
Orchid
An LLM-driven narrative authoring system validated with 100+ users and later published at ACM Creativity and Cognition, showing my ability to turn research into a usable interactive system.
Multi-Agent Systems + RAG • 2025
NarrativeHive
A multi-agent LLM social simulation with persistent memory and local-model support, built to explore coherence, retrieval, and agent behavior over long-running interactions.
Speech AI + Healthcare • 2025
HK-GenSpeech
A speech-based cognitive-screening prototype built with Whisper, PyTorch, torchaudio, ffmpeg, and React Native to turn generative scene prompts into clinically useful speech elicitation.
Experience
Independent AI Consultant at 2084 Futures Limited
Edge-AI feasibility for embedded camera vision: compared Tiny/MobileViT + INT8 against Coral/Hailo/IMX500 under real-time surveillance constraints; delivered build-vs-buy guidance and deployment recommendations.
AI Systems Engineer at dRoW Limited
Built private OCR/LLM evaluation pipelines with vLLM, Ollama, and MLflow to benchmark 100+ vendor/invoice templates without moving sensitive data off-prem. Added regression checks for schema drift and unstable outputs; delivered the evaluation report adopted for production vendor selection.
Applied AI Researcher & Technical Lead at HKUST (HLTC/XRIM Labs)
Led applied AI R&D across narrative systems, multimodal interfaces, and speech tooling. Built NarrativeHive, Orchid, SeaSense, and HK-GenSpeech; mentored 30+ student engineers per semester and translated research goals into stakeholder-facing demos and shippable technical scopes.
Software Engineer at Arskom Group
Migrated monolithic satellite billing platform to modular Python/Django architecture, improving billing accuracy by ~40%. Used Jenkins + PyTest to reduce system error rates by ~75% and stabilized weekly releases.
Earlier Career Foundation
Quantitative Linguistics Analyst, METU
Built annotation pipelines and statistical reporting for publication-oriented language research, including ANOVA-based analysis and publication-ready reporting.
Freelance Web Developer
Delivered client websites end-to-end, building the product instinct and full-stack execution discipline that later carried into applied AI work.
Publications
7 peer-reviewed publications spanning NLP, multimodal systems, speech processing, and AI evaluation. These track the same arc as the projects: turning research problems into measured, reproducible results.
Morphological Segmentation and Bayesian Models for Turkish Language Processing
Kumyol, Serkan
Modeling morpheme triplets with a three-level hierarchical Dirichlet process
Kumyol, Serkan and Can, Burcu
Allomorphs and binary transitions reduce sparsity in turkish semi-supervised morphological processing
Can, Burcu and Kumyol, Serkan and Bozsahin, Cem
HK-GenSpeech: A Generative AI Scene Creation Framework for Speech Based Cognitive Assessment
Yong, Vi Jun Sean and Kumyol, Serkan and Low, Pau Le Lisa and Leung, Suk Wai Winnie and Braud, Tristan
Myokey: Surface electromyography and inertial motion sensing-based text entry in ar
Kwon, Young D and Shatilov, Kirill A and Lee, Lik-Hang and Kumyol, Serkan and Lam, Kit-Yung and Yau, Yui-Pan and Hui, Pan
Efficient Bilingual Generalization from Neural Transduction Grammar Induction
Yan, Yuchen and Wu, Dekai and Kumyol, Serkan
Orchid: A Creative Approach for Authoring LLM-Driven Interactive Narratives
Wu, Zhen and Kumyol, Serkan and Wong, Shing Yin and Hu, Xiaozhu and Tong, Xin and Braud, Tristan
Technical Skills
Languages
- Python
- C++
- JavaScript/TypeScript
- SQL
- Bash
ML/LLM
- PyTorch
- TensorFlow
- DyNet
- Hugging Face
- scikit-learn
- LangChain/LangGraph
- RAG Workflows
- vLLM
- Ollama
Data & Web
- React
- Next.js
- Node.js
- PostgreSQL/MySQL
- Vector Search
Infra
- Docker
- GitHub Actions
- Jenkins
- AWS (EC2, S3, Lambda, SageMaker)
- GCP
- Azure
- Linux/Unix
Other Tools
- MLflow
- Kafka
- Whisper
- torchaudio
- ffmpeg
- Vision Transformers
- INT8 Quantization
- Jupyter
- Benchmark design
- Regression testing
Education
Training that crosses boundaries: machine learning, cognitive science, and education. This shaped how I build systems and how I explain them—whether to students, collaborators, or clients trying to make build-vs-buy decisions.
Ph.D. in Computer Science (Machine Translation and AI Systems)
Hong Kong University of Science and Technology (HKUST)
Research at the intersection of machine learning, narrative systems, multimodal interaction, and MLOps. Focus: LLMs, multi-agent systems, speech processing, and UI/UX for interactive AI. Concurrent applied AI project delivery throughout the PhD.
M.Sc. in Cognitive Science (Artificial Intelligence and NLP)
Middle East Technical University (METU)
Studied language, learning, and perception, with thesis work on unsupervised morphological learning in Turkish using probabilistic models (Hierarchical Dirichlet Process). Focus on Bayesian methods and NLP.
B.Sc. in Computer Systems & Education
Gazi University
Built foundation in computing while training in how to teach technical material clearly, structurally, and at the right level for the audience.
Credentials & Leadership
AWS
AWS Certified AI Practitioner
Formal cloud-AI credential aligned with production deployment and model-infrastructure decision making.
2024
President, RoastMasters HK Toastmasters Club
Led a 50+ member club, organized evaluations and events, and sharpened stakeholder communication under live-feedback conditions.
2023
3rd Place, Best Speech Evaluation
External signal for clear, structured feedback and executive-facing communication.
Get in touch
Want to collaborate or talk AI?
I'm open to research collaborations, applied AI consulting, and conversations about evaluation, deployment, and the craft of building systems that work.