I build AI systems end to end: data, models, evaluation, inference, and the software people use them through. For more than five years I have taken computer vision, Document AI and LLM applications from proof of concept to production, and I build independent desktop products on FluentQt, the Python UI framework I created.

• LLMs, RAG and agents. Agentic RAG, hybrid and graph retrieval, reranking, structured output, and local inference with Ollama, vLLM, SGLang and llama.cpp.
• Document AI. OCR, extraction with vision-language models, field localisation, validation and human review.
• Computer vision and industrial AI. Detection, segmentation, tracking and anomaly detection, connected to cameras, PLCs, Modbus, MQTT and edge devices.
• Application engineering. Python desktop software (PyQt5, FluentQt), FastAPI services, ONNX deployment and standalone builds.

The source of these projects is private. Each card opens a public showcase with documentation and screenshots of the real application.

Veyra is a commercial machine-vision and AI inspection platform, developed since 2022. It has 72 vision and AI tools, camera and PLC stations, OK/NG evidence and labelling, and Ultralytics and Anomalib training with parity-checked ONNX exports (55 of 60 models verified end to end on CPU).

FluentQt is a Fluent Design UI framework for PyQt5 in Community and Pro editions, maintained since 2022. Community has 177 widgets in 12 families, theming, frameless windows, six languages and SQLite persistence. Pro adds data grids, charts, a node editor and computer-vision UI. The other seven featured applications are built on it.

ComicTranslator translates manga and comics locally. It detects bubbles, runs multilingual OCR, translates with a GGUF LLM through llama.cpp, then inpaints and letters each page for review.
_{Pages shown: Pepper&Carrot episode 3 by David Revoy, CC BY 4.0.}

Vietnamese Legal Intelligence is a legal research workstation. It parses 3,765 legal documents into more than 100,000 provisions and stores 120,000 citations in a Neo4j knowledge graph. Hybrid retrieval combines BM25, multilingual E5 and rank fusion. Answers come from a local LLM, and each one is cited, checked against its sources, or declined.

Fluent Data Studio answers plain-language questions about data in a way you can check. It copies files, databases and web APIs into a local DuckDB workspace, splits a request into requirements and runs each as a validated plan of typed operations, so every result shows what it answers and the SQL behind it. Charts stay editable, and kept results go onto a report.

AI Annotation Workbench labels images and video for computer vision: boxes, rotated boxes, polygons, keypoints and tags. 16 local ONNX models across 11 task families propose annotations on a laptop CPU, and nothing is exported until a person accepts it. It evaluates models with each task's metric and exports YOLO, COCO, DOTA, MOT and PaddleOCR.

ChessForge is a 3D chess game and chess workstation. Its own OpenGL renderer draws five original sculpted piece sets, six boards and five arenas that differ in shape, not just colour, with original sound and music, a twelve-mission campaign, a cosmetic collection and a camera that glides between two players on one computer. It also reviews games with Stockfish, trains with puzzles and opening drills, and keeps a game library; its AI Lab sets up a private PyTorch runtime from one button (processor or NVIDIA GPU), trains a policy/value network in a separate process, runs MCTS self-play and lets you play against your own models.

Audio Story produces narrated audio stories, Vietnamese first. The author plans, casts and writes with an OpenAI-compatible LLM whose proposals they accept or reject; local text-to-speech (VieNeu-TTS, Piper) voices each line, and episodes are mixed to BS.1770 loudness and exported as a package ready for YouTube.

Independent re-implementations of systems I built professionally. They use synthetic data only and contain no production code or customer data.

DocFlow IDP turns a mixed PDF packet of shipping documents into validated structured data. It classifies and segments the pages, extracts schema fields, locates each value with a separate bounding-box pass, checks business rules (totals, dates, ISO 6346 container check digits) and routes doubtful values to a human review queue with an audit trail. Every model call goes through one backend contract, so an offline baseline and VLM backends are interchangeable.

T-Shirt Quality Inspection checks a printed T-shirt from one top-down photograph: QR job identification, millimetre calibration, garment segmentation and landmarks, size measurement, print alignment to the artwork, and SSIM and ΔE difference analysis. Seven checks decide PASS or FAIL, and each one states its value, limit and reason. The re-implementation uses deterministic, weight-free computer vision on synthetic photographs and is built with FluentQt.

Date Code Inspector reads the production and expiry dates printed on cartons, bottles, pouches and cans, character by character, then validates them against the product's rules and decides GOOD or NG with the failed rule named. On its synthetic demo set, 22 of 22 verdicts match the ground truth and 597 of 600 characters are read correctly. Built with FluentQt.

More re-implementations

Research implementations. Results and write-ups of detection models I implemented: YOLO (v1 to v11 and YOLOX), RYOLO (oriented detection), YOLO-Seg (instance segmentation), EAST and CRAFT (scene-text detection).

M.Sc. and B.Eng. in Electrical and Electronics Engineering (Automation), Ho Chi Minh City University of Technology.

pntrungbk15@gmail.com · Ho Chi Minh City, Vietnam