VAAMBARTSUMOV
AI SYSTEMS

KL / MY · 2026

PORTFOLIO / 01–05
DATA · MODEL · SYSTEM

ML / AI ENGINEER

Make the data
tell the truth.

I build open AI systems end to end — from trustworthy data and evaluated speech models to agents, robotics, and deployment.

Open to internships and junior ML/AI engineering roles · Remote-first · US, Gulf, and selected APAC opportunities.
Explore the work Start a conversation
Official portrait of Vyacheslav Ambartsumov
01

Systems builder.
Evidence first.

Kuala Lumpur, Malaysia

How I work

Build the dataset.
Measure the model.
Ship the system.

01Work from human signal, not assumed labels.

02Keep trade-offs and error profiles inspectable.

03Connect models to the boundary where they must work.

Selected work

Proof before
promises.

Three systems that show the complete loop: finding signal, making a model earn trust, and carrying that work into an actual system.

01

Selected system

Low-resource speech recognition, built from the data layer upward.

Turkmen ASR — Open Data & Evaluation Stack

A public, reproducible Turkmen ASR stack that confronts an overlooked evaluation problem: machine-generated references can hide model error. The work connects dataset curation, human evaluation, transcript correction, LoRA fine-tuning, and transparent error analysis.

HELD-OUT EVALUATION
124.61%24.44%
Fixed clips, decoder and normalization. 80% relative WER reduction.
Evidence
Fine-tuned WER
24.44%fixed 300-clip held-out evaluation
Zero-shot baseline
124.61%same clips, decoder, and normalization
Relative WER reduction
80%vs. zero-shot Whisper Large v3 Turbo
Training corpus
251 hnatural spoken Turkmen with LoRA fine-tuning
PythonHugging Face TransformersWhisper Large v3 TurboLoRAASR evaluationDataset curation
GitHub data reviewFine-tuned modelHuman evaluation set
02

Selected system

A physical AI system that connects perception, language, memory, and movement.

Autonomous Robot Kesha

A low-cost wheeled home robot built around ESP32-CAM hardware and a Python brain. Kesha integrates local language models, object detection, speech interfaces, sensor-driven navigation, memory, scheduling, and actuator control.

Integrated camera, microphones, motors, distance sensors, and a language-model control layer.

Built a documented REST boundary for brain, speech, vision, and system status.

Iterated from basic navigation to a multi-modal autonomous agent architecture.

Evidence
Hardware cost
~$53core components before shipping
Control stack
ESP32-CAM ↔ FastAPIWi-Fi hardware/software boundary
Perception
Vision + speechYOLOv8n, Faster-Whisper, Piper TTS
Local intelligence
Ollamalocal LLM with cloud fallback
PythonFastAPIESP32-CAMYOLOv8Faster-WhisperPiper TTS
GitHub repository
03

Selected system

A research platform for measuring whether critique improves engineering problem solving.

Self-Improving AI Solver

A full-stack experimental environment that compares baseline LLM responses with a separate critic-and-revision path across engineering problems. The platform persists runs, exposes analysis, and makes the comparison inspectable rather than anecdotal.

Implemented baseline and critic workflows as separate experimental conditions.

Built a dashboard and REST API for problems, runs, analysis, and administration.

Made improvement, accuracy, and error-detection data available for inspection.

Evidence
Architecture
Baseline → Critic → Revisiontwo-stage comparative loop
Problem set
32 problemsacross four engineering domains
Interface
Full stackReact dashboard, FastAPI API, persisted experiments
Analysis
Live metricscomparison and error-detection views
PythonFastAPIReactSQLAlchemySymPyDeepSeek
GitHub repository

Operating range

From raw context
to working logic.

01

Data & evaluation

ASR corpora, human references, transcript correction, WER/CER and error analysis.

02

Model adaptation

Whisper, LoRA fine-tuning, reproducible training and comparison.

03

Agent systems

LLM routing, tools, memory, task planning and controlled execution.

04

Physical AI

ESP32-CAM, YOLOv8, voice, sensors and autonomy boundaries.

05

Applied stack

Python, FastAPI, SQLAlchemy, Docker, React and APIs.

06

Systems discipline

Tests, clear interfaces, local-first constraints and written trade-offs.

Supporting work

Other systems
under load.

04

PDS-Ultimate

A documented multi-tool assistant architecture with voice and text control, offline speech recognition, task routing, containerized deployment, and an extensive test suite.

Python · DeepSeek · Telegram

Source
05

SentinelForge

A local, LLM-driven security-audit prototype for authorized, isolated lab environments. It connects a FastAPI agent core, attack-graph visualization, local inference, and controlled tools within a documented hackathon project.

Python · FastAPI · vLLM

Source

Path

Learn fast.
Own the work.

2026–27

Monash University Foundation Year

Building the academic base for a planned progression into AI Engineering at Monash University Malaysia.

2024–26

Dgtl · Ashgabat

Built a CRM workflow around company-specific operations, with AI-agent assistants to reduce day-to-day friction.

100 / 100

ITEA Programming Diploma

International programming coursework completed with a red diploma. Russian: native. English: B2.

Next system

Let’s build something
that matters.

I am open to internships and junior ML/AI engineering roles where careful thinking, fast learning and direct ownership are valuable from the first day.

Emailslavikam37@gmail.comTelegram@j40I7GitHubambartsumovhfHugging FaceSlavik2009