Portfolio2026
Kaiyrly Mukhametkarim
AI engineer. I design, train and ship production AI systems. LLMs, RAG, speech and AI agents, for Kazakh and Russian speaking users.
Computer Science graduate of UNIST in South Korea. Since 2023 I have delivered 40+ client projects. Today my AI systems run in production for pharma teams, medical researchers, an enterprise and local businesses.
Education
UNIST, South Korea
BSc Computer Science and EngineeringUlsan National Institute of Science and Technology, 2020 to 2024
Research and teaching at UNISTSecurity lab intern, ART Lab engineer, Web Technologies instructor
ACM ICPC Seoul RegionalsTop 25, 2022
Informatics olympiadsZhautykov, Eurasian and National medals, 2019 to 2020
Now
AI engineer, iKomek and freelance
MDWRT Medical Research CopilotUsed by pharma teams, including Pfizer
RA-EDC clinical registry1,239 medical documents structured by AI
iKomek knowledge base and AI chatSmart search and AI chat on its own servers
On-prem legal assistantOpen model chosen and tested for 16 users on one box
Curriculum vitaeKaiyrly Mukhametkarim
Experience
Apr 2025 – now
AI engineer · iKomek, KazakhstanBuilt the knowledge base, smart search and AI chat on the company's own servers. Trained Kazakh speech models, built call quality scoring, made answers about 6× faster under load.
2023 – now
Freelance AI engineer · Remote40+ client projects in pharma, medical research, EdTech, legal and fintech. Built and run MDWRT, RA-EDC, an AI training platform for pharma reps and an on-prem legal assistant.
Apr – Aug 2024
CTO · Limitlessmind, remoteLed AI automation work for large corporate clients. Closed $50K in contracts in five months.
Jun 2023 – Jan 2024
Full-stack developer · LATOKEN, remoteBuilt marketplace features in React and Django that doubled user engagement. Fixed critical production bugs.
Sep 2023 – Jan 2024
Assistant instructor · UNIST, South KoreaTaught Web Technologies to 30 students. Ran lectures and supervised projects.
Sep 2022 – Jun 2023
Software engineer · ART Lab, UNISTPython backends on AWS Lambda and EC2. Ported C++ to WebAssembly microservices. Trained models for algorithm matching.
Dec 2022 – Mar 2023
Software engineer · OnSquare, SeoulBuilt an audio editor from scratch with D3.js visualization. Doubled audio effects performance.
Jun – Dec 2021
Research intern · Computer Systems Security Lab, UNISTBinary similarity detection at 90% accuracy without source code. Located vulnerable functions in patched binaries at 75% accuracy.
Education
2020 – 2024
BSc Computer Science and Engineering · Ulsan National Institute of Science and Technology (UNIST), South KoreaData structures, AI, networks, databases, software engineering, compilers, interpreters.
Skills
- AI and LLM
- Qwen, LangChain, LlamaIndex, RAG, GraphRAG, tool calling, LLM-as-judge evals, full and LoRA fine-tuning, reinforcement learning
- Model serving
- vLLM, llama.cpp, quantization (GGUF, NVFP4), speculative decoding, load testing
- Search and data
- Postgres, pgvector, Qdrant, OpenSearch, Pinecone, Redis
- Backend
- Python, Django, DRF, FastAPI, Celery, aiohttp, Node.js
- Frontend
- TypeScript, React 19, Next.js 15, Expo and React Native, Tailwind
- Voice and video
- LiveKit, Pipecat, Asterisk SIP, ASR evaluation, Higgs Audio, LTX and Wan video models, ComfyUI
- Infra
- Docker, Kubernetes, Linux servers, on-prem GPU nodes, Modal, AWS, CI, Pytest, Playwright
Awards and languages
- ACM ICPC Seoul Regionals, Top 25 · 2022
- Zhautykov Olympiad in Informatics, Bronze · 2020
- Eurasian Olympiad in Informatics, Bronze, Silver · 2019–20
- National Olympiad in Informatics, Bronze, Silver · 2019–20
- Languages, Kazakh, Russian, English
Case study · Pharma AI
MDWRT Medical Research Copilot
01
An AI workspace for pharma medical teams at mdwrt.app. They use it to find research, read and compare documents, and turn the evidence into slides, summaries and briefings. Tasks that took hours now take minutes.
22AI tools in one workspace
500+research papers reviewed for one presentation
Minutesinstead of hours for routine research
What the platform does
Research
Search medical literature, get alerts on new papers, write literature reviews and research summaries, and draft regulatory documents.
Documents
Compare documents side by side, summarize them, pull out clinical trial results, translate medical texts and check facts.
Content
Create evidence-based slide decks, infographics, data stories and audio versions of materials for field teams.
Team knowledge
A shared wiki, project spaces and a bot that answers questions from the team's own materials.
What I built
- Designed the AI pipelines behind all 22 tools. Search, reading, comparison and writing, each tuned for medical texts.
- Multi-step AI agents that go from a research question to finished slides.
- The AI reads the full research paper, not just the abstract. Each file is checked to be the right paper.
- Pick the right AI model for each task, measured on real documents for quality and cost.
DjangoReactCeleryRedisPostgresLangChainLlamaIndexOpenAIAnthropicPubMed API
MDWRT · Evidence Deck and Deck Audit
Build the deck, then prove every claim
01
Slide decks for doctors must be correct down to every number. I built two tools for this. Evidence Deck researches a drug and builds an editable PowerPoint deck. Deck Audit checks any deck, claim by claim, against the papers it cites.
Create
Evidence Deck
It plans the deck, collects the research, fills the gaps, and builds the slides after approval. Every number shows whether it was checked against the full paper.
90slides delivered to a global pharma company
206full research papers behind them
10charts from papers rebuilt as editable PowerPoint charts
96slides laid out in about 30 minutes
- Research runs on 10 topics at once, from effectiveness and safety to treatment guidelines.
- Graphs from papers become real PowerPoint charts the client can edit.
- 391 references link straight to the source paper.
Check
Deck Audit
Upload a PowerPoint or PDF. It finds every claim, opens the cited paper, and marks each claim as supported, partly supported or not supported.
8 of 9errors found that an expert reviewer had found
~$0.30per deck, about 14 minutes
62%cheaper in the new version
2×faster in the new version
- Each finding shows the exact quote from the paper.
- Reads charts and tables on slides, not only the text.
- Reports in Word and web formats, ready for SharePoint.
- Tested on real decks before every release.
python-pptxPDF parsingChart digitizingOpenAIVision modelsCeleryPostgres
Case study · Clinical AI
RA-EDC clinical research registry
02
A research database for rheumatoid arthritis. Doctors' notes, hospital discharge papers and lab reports arrive as scans and PDFs in Russian. The system reads them, turns them into clean structured data, and shows which patients are hard to treat.
1,239medical documents turned into structured data
50+types of clinical data, from visits and lab results to treatments and side effects
11months of continuous development
What I built
- A complete patient record. Visits, lab results, scans, treatments, infections, vaccinations, side effects and treatment plans.
- The AI reads each part of a document separately, then checks its own output and fixes mistakes automatically.
- Disease activity scores and a hard-to-treat check, each linked to the original text.
- Processes hundreds of documents at once and tracks the cost of each one.
- Results are compared with a hand-checked sample before every change.
From document to data
Uploadscans, PDF and Word in Russian
Readtext recognition, match to patient
Extractfacts for each section
Checkverify and fix
Scoredisease activity, hard to treat
Databasepatient list, statistics, export
Also built. A data collection system for a breast cancer drug study. A tool that reads survival curves from research papers in seconds, instead of 10 to 15 minutes by hand.
DjangoCeleryRedisOpenAIAnthropicTesseract OCR
Case study · Self-hosted AI
iKomek knowledge base and AI chat
03
iKomek answers citizens' questions about city services in Kazakh and Russian. I built and trained the AI system behind it. A RAG search made for its spreadsheets and documents, a large language model running on the company's own modest servers, and trained Kazakh speech models.
35Blarge AI model running on one mid-range graphics card
6×faster answers under heavy load
16 / 16people chatting at once with no errors
RAG search built for this data
- Understands spreadsheets. Each row keeps the context of its table, so the chat can answer “when is my street served”.
- Finds answers by meaning and by exact words.
- Understands follow-up questions in Kazakh.
- Every answer cites its source.
Self-hosted LLM on modest hardware
- Each server has a 12 GB graphics card and 16 GB of memory.
- A large 35B model still runs on it, fast enough for live chat.
- Two servers share the load. Cloud AI is only a backup.
- Found the bottleneck and made answers about 6× faster.
Speech models and call quality scoring
- Trained Kazakh speech-to-text and text-to-speech models.
- Tested 10+ speech models on real calls. Found one that handles mixed Kazakh and Russian in a single pass.
- Operator and caller are split by channel, then an AI judge scores each call on a 12-point rubric.
Knowledge base for the AI
- Upload PDF, Word and Excel files. They are parsed, split and indexed for AI search automatically.
- Handles very large files, even FAQ workbooks with millions of characters.
- A spreadsheet editor inside the app. Edit cells, rows and columns like in Excel, and the AI sees the change.
DjangoPostgresQdrantOpenSearchllama.cppQwenWhisperLoRA
Case study · On-prem LLM
Legal assistant on one on-prem box
04
A client needed answers about Kazakhstan cybersecurity and personal data law, in Russian and Kazakh, on their own hardware with no cloud. I built the knowledge base, chose the open model and proved the setup on the real machine.
80 of 84questions answered correctly on the real box, with 0 invented answers
16people at once, each at 13.7 tokens per second
~$40total cost of the whole evaluation
What I built
- A legal knowledge base with 180 topics. Each norm keeps its source, its cross-references and whether it is still in force.
- The model navigates topics instead of plain search. Navigation reached 100% of the right norms. Hybrid search reached 81%.
- Tested about 10 open models on rented GPUs, graded automatically against checked answers.
- The winner is a 35B Qwen model in 4-bit NVFP4 on vLLM. Russian 91%, Kazakh 81 to 83%.
- Kazakh was at 15% because the model kept answering in Russian. A Kazakh system prompt fixed it.
- Speculative decoding made the background model 2.1× faster.
- A background model running nonstop halved live speed, so background jobs run only when load is low.
How the model was chosen
Questionsclient questions with checked answers
Knowledge baselaws split into 180 topics
Bake-offabout 10 models, same questions
Real boxload test with 16 users
Verdictmodel, settings and background plan
vLLMQwenNVFP4Speculative decodingTopic navigationRAG baselineModal GPUsVast.ai
Case study · AI tutoring
AI Training Platform for pharma reps
05
A platform that trains pharmaceutical sales reps. Managers upload product knowledge and set goals. Reps learn with an AI tutor by text or voice, then take tests graded against those goals. The platform shows who is ready and who needs help.
3AI roles. Tutor, examiner and supervisor
3,400+automated tests keep it reliable
< 3months from specification to a working system
What I built
- An AI tutor that teaches only from the company's own materials.
- An AI examiner that writes and grades tests based on the learning goals.
- A manager dashboard that shows progress for each person and topic.
- The AI sticks to company materials and refuses questions outside them.
- Automatic AI evaluations check grounding, grading and search quality before every update.
- Built for compliance. Actions are logged and personal data can be fully erased.
- Voice lessons are designed and ready to switch on.
Learning loop
Manageruploads knowledge, sets goals
Tutorlessons by text or voice
Examinertests and grading
Reviewrepeats what was forgotten
Dashboardwho is ready, who needs help
FastAPIPostgrespgvectorCeleryNext.jsLiveKitAnthropicOpenAIElevenLabsDocker
Case study · Generative AI
Kazakh voice and video AI
06
There are very few AI voice and video tools for Kazakh. I train AI models on a specific person, their voice and their face, and use them to create talking videos at a fraction of studio cost.
30 minof recordings is enough to clone a Kazakh voice
$0.70per minute of realistic AI avatar video
1 photois enough to make a person talk, with accurate lip-sync
What I built
- Trained Kazakh voice models from about 30 minutes of a person's speech.
- The fine-tuned voice reads dates, times and amounts correctly, where base models failed.
- Compared several voice engines and picked the winner by listening, not only by numbers.
- The voice stays consistent across long lessons.
- Fine-tuned the LTX-2.3 video model (22B parameters) on each person's real footage, so the avatar really looks and moves like that person.
- Also built lip-sync with Wan 2.2, InfiniteTalk and MiniMax H3.
- Long videos with no visible cuts, and audio that stays in sync.
From recording to video
Record30 minutes of voice and short video clips
Preparetranscribe, cut into sentences, crop faces
Traina voice model and an LTX-2.3 video model for the person
Generatescript to voice to talking video
Finishsharpen and join into one video
Higgs AudioLoRALTX-2.3Wan 2.2musubi-tunerComfyUIModal GPUsSeedVR2Soniox
Case study · EdTech
Natije learning platform
07
A math learning platform in Kazakh and Russian for grades 5 and 6, live at natije.online. Lessons, practice, diagnostics and parent reports, plus a pipeline that turns lesson plans into narrated videos checked by AI judges.
6,000+automated checks behind the latest release
51grade 6 guides in two languages, reviewed and playable
3AI judges for each video lesson. Pedagogy, narration and visuals
What I built
- A diagnostic test places each child, then a daily plan follows.
- Lessons and practice in Kazakh and Russian for grades 5 and 6.
- Parents get reports and weekly recaps, with consent and data erasure built in.
- Each video lesson passes a mechanical check and then the AI judges. Failed lessons are fixed and checked again, up to three times.
- Quality comes from the checks, not from which model writes the lesson. A lesson written by a cheaper model scored 12 of 12 on pedagogy.
- Every release is tested on phone screens in both languages before it goes live.
From lesson plan to video
Planlesson plan and scenes
Checkpacing, terms and timing rules
Judgespedagogy, narration, visuals
Fixup to three rounds
Voice and renderKazakh narration and video
Next.js 15PostgresDrizzlePlaywrightLLM judgesKazakh TTS
08 · Agents
AI agents for business
AI that talks to customers where they already are. A sales agent answers leads in WhatsApp and Instagram. A voice agent takes and makes calls in Kazakh and Russian. Tutors and bots answer only from checked material.
WhatsApp and Instagram sales agent Live
Answers questions from the business's own information, finds serious buyers and hands them to a person when needed. It fully follows the sales script, types like a human and converts better than human sales managers. Every prospect gets a personal demo.
Phone voice agents Built
Take and make calls through a local phone operator and book meetings into Google Calendar. Built on LiveKit and Pipecat. They speak Kazakh with a custom voice model and hand the call to a person when needed.
Course tutor from video lessons Client project
62 Kazakh course videos were transcribed and turned into topics, rules and a FAQ. The tutor answers only from that material, in Kazakh and Russian.
Payment receipt check in chat Live
A Telegram bot reads a Kaspi receipt as a PDF, a screenshot or a link, verifies it and then sends the materials. Nobody checks receipts by hand.
PythonWhatsApp APIInstagram APILiveKitPipecatAsteriskGoogle Calendar
Contact
Let's build something that works in production
I take AI products from idea to a running system. Research, models, backend, frontend and servers.