Minh Khang Tran

Machine Learning Engineer  ·  Computer Vision & LLM Systems

Paris, France  ·  Open to Machine Learning and Research Engineer roles
LinkedIn GitHub khangtm99@gmail.com

Work Experience

Master's Thesis Researcher Jan 2026 – Sep 2026
University of Amsterdam  ·  with 3DUniversum  ·  Amsterdam, Netherlands
  • Developed Cross-Illumination Albedo Invariance (CIAI), a training strategy that requires the albedo predicted from two photographs of one scene under different lighting to agree, and the shading to explain the brightness change between them.
  • Cut the luminance variation of a material across lighting by 39% (0.259 → 0.157 on the Multi-Illumination Dataset) in a matched ablation, and lowered albedo error on unseen coloured-light renderings (ARAP) by 7–14%. Rendered 5,056 coloured-light image pairs with exact albedo for the next stage.
  • Trained about 70× fewer parameters than the diffusion baselines, with about 5× faster inference. The method does not yet make colour stable, because its explanation loss uses luminance only; the next model adds a separate shading-chroma factor.
  • Supervised by Prof. Theo Gevers and Dr. Sezer Karaoglu (UvA), with Prof. Seyed Ali Amirshahi and Dr. Luis Gomez Robledo (COSI). Submitted July 2026, defended September 2026 at NTNU.
With the 3DUniversum team, Amsterdam
With the 3DUniversum team, Amsterdam
Machine Learning Intern Jun 2025 – Aug 2025
DXOMARK  ·  Paris, France  ·  Supervisor: Dr. Sira Ferradans, AI Scientific Director
Vision-language models for portrait image quality
  • Adapted the vision-language model DeQA-Score from one overall quality score to per-attribute portrait scores (noise, texture, contrast) for camera portrait-mode benchmarking.
  • Converted Just-Objectionable-Difference (JOD) annotations, a psychophysical scale built from pairwise expert comparisons, into a MOS-compatible training scheme; evaluated on KonIQ, SPAQ, KADID and internal portrait data.
  • Compared regression-head, vision-only, and LoRA fine-tuning alongside prompt and cropping strategies, matching expert ratings with SRCC 0.91/0.93 and PLCC 0.91/0.90 (noise/texture). Ablations showed the vision encoder drove most of the gain.
  • Co-authored a paper that won the Best Student Paper Award at IS&T Electronic Imaging 2026.
The DXOMARK team, Paris
The DXOMARK team, Paris
AI Engineer Oct 2023 – Dec 2025
Vitex Software JSC  ·  Hanoi, Vietnam (Remote)
Full-time until Jul 2024, then part-time remote during the MSc.
  • Led development of PlanZ's multi-agent assistant, handling hundreds of requests per day for scheduling, supplier discovery, and natural-language questions over live event data (LangGraph, tool calling, Qdrant-based RAG, text-to-SQL, FastAPI, PostgreSQL, deployed on AWS).
  • Routed routine steps to a smaller, cheaper model and kept the larger model for complex planning, to control inference cost.
  • Set up agent evaluation with LangSmith traces and fixed failures through prompt engineering, structured outputs, retries with error feedback, and schema-grounded SQL, cutting the failure rate from ~15% to ~5%.
VLM OCR pipeline — handwritten Japanese vaccination forms
  • Built an event-driven pipeline that turns handwritten Japanese vaccination forms (予診票) into 14 structured fields, replacing manual data entry at 2,000 images/hour with ≥92% field-level accuracy and ≤5.6% CER, using a self-hosted Qwen3-VL-8B on vLLM.
  • Fine-tuned the model with LoRA, quantization, and visual grounding on human-verified corrections of its own flagged mistakes.
  • Built in reliability: tiered RabbitMQ retries, a Redis-backed circuit breaker, idempotent updates, and HMAC-signed webhooks.
  • Built an internal tool HR uses to shortlist the top 10 candidates per role: GPT-4o-mini extracts structured profiles from CVs and job descriptions, embeddings in PostgreSQL/pgvector rank them, and an optional LLM rerank explains the matches (FastAPI, Docker, React).
With the Vitex team
With the Vitex team
AI Engineer May 2023 – Aug 2023
Viettel High Tech (VHT), Viettel Group  ·  Hanoi, Vietnam  ·  Supervisor: BangLV
Video surveillance system
  • Deployed face-recognition models (detector + ArcFace, exported to ONNX) on Kubernetes and NVIDIA Triton Inference Server, increasing inference throughput by ~3× through dynamic batching and concurrent execution.
  • Enabled zero-downtime model updates through runtime model load/unload and service discovery.
  • Refactored the controller block for scalability and built a service-discovery server that tracks available services and model metadata.
With the Viettel High Tech team
With the Viettel High Tech team
AI Engineer Intern Jun 2021 – Dec 2021
Pionero JSC  ·  Hanoi, Vietnam  ·  Supervisor: Huy Van
Facial recognition attendance system
  • Built a facial-recognition attendance system on existing office cameras (YOLOv5 + FaceNet + softmax classifier) in a 10-person Scrum team, shipped through CI/CD, cutting check-in waiting time by 70%.
  • On an internal set of 10,188 annotated frames (7,333 face boxes), reached F1 0.90 for face detection and F1 0.92 for recognition, running end-to-end at 30–40 FPS on 512p input.
  • Fine-tuned detection and recognition for masked faces during COVID-19.
The Pionero team, Hanoi
The Pionero team, Hanoi

Earlier Research

Student Research Assistant Oct 2021 – Mar 2022
International Research Institute MICA (HUST)  ·  Mentor: Dr. Thanh-Hai Tran
Polyp segmentation (mentored project)
  • Prepared and quality-controlled polyp datasets (Kvasir-SEG, CVC-ClinicDB, ETIS-Larib).
  • Retrained TransFuse with Swin Transformer blocks instead of ViT, improving mDice from 0.733 → 0.77 and mIoU from 0.659 → 0.70 on ETIS.

Research projects (HyRES, TextMAE, hyperspectral inspection with Dispelix) and coursework reports are on the Research page.

Education

MSc Computational Colour and Spectral Imaging Aug 2024 – Sep 2026
Erasmus Mundus Joint Master  ·  Norway → France → Finland
NTNU Université Jean Monnet University of Eastern Finland
  • Admitted 1st of 198 applicants on a full Erasmus Mundus Consortium Scholarship. Grade: Very Good.
  • Relevant coursework: Advanced Image Processing, Advanced Deep Learning, 3D Models in Computer Vision, Light-Matter Interaction and Materials Appearance, Advanced Spectral Imaging Devices, Colour Science Laboratory, From Statistics to Data Mining, Pattern Recognition, Eye Tracking.
  • Thesis at the University of Amsterdam with industry partner 3DUniversum (see above).
COSI cohort at NTNU, Gjøvik
COSI cohort at NTNU, Gjøvik
Université Jean Monnet, Saint-Étienne
Université Jean Monnet, Saint-Étienne
UEF Photonics Center, Joensuu
UEF Photonics Center, Joensuu
Graduation day, NTNU
Graduation day, NTNU
BSc Computer Science Jan 2019 – Dec 2022
State University of New York at Oswego  ·  Oswego, NY, USA
  • GPA 3.81 / 4.0 (major 3.82), Summa Cum Laude; President's List (Spring 2019, Spring 2021, Fall 2022) and Dean's List.
  • Destination Oswego Scholarship and Merit Scholarship.
Graduation, SUNY Oswego
Graduation, SUNY Oswego
  • 2026Best Student Paper Award, IS&T Electronic Imaging 2026, for Adapting DeQA-Score for Attribute-Specific Portrait Quality Assessment (Yujin Cho, Minh Khang Tran, Benoit Pochon, Jean-Michel Morel, Gabriele Facciolo, Sira Ferradans)
  • 2024Erasmus Mundus Consortium Scholarship, full scholarship, ranked 1st of 198 applicants
  • 2022Summa Cum Laude, SUNY Oswego; President's List ×3, Dean's List ×3
  • 2019Destination Oswego Scholarship and Merit Scholarship, SUNY Oswego
Best Student Paper Award certificate, IS&T Electronic Imaging 2026
The award certificate, listing the co-authors
Programming
Python SQL Java Go
Machine Learning & Vision
PyTorch TensorFlow Scikit-learn OpenCV Transformers / VLMs LoRA Diffusion Models CNN GAN / VAE Image Compression
LLM Applications
LangGraph LangChain LangSmith RAG Tool Calling Text-to-SQL Qdrant pgvector OpenAI API MCP
Serving & Infrastructure
Docker Kubernetes Triton Server vLLM ONNX FastAPI Flask PostgreSQL Redis RabbitMQ MinIO MongoDB MySQL AWS Git
Practices
Agile / Scrum Jira Unit Testing CI/CD Agent Evaluation Prompt Engineering
AI Development Tools
Claude Code Codex
Languages
Vietnamese (native) English (C1) French (learning)
EM Sharing 2024 organisers and participants
EM Sharing 2024, which I helped organise
Erasmus Mundus scholars from Vietnam at the EU Delegation
With Erasmus Mundus scholars from Vietnam at the EU Delegation

The trait I value most in an engineer is teaching yourself. AI was not part of my curriculum when I started at SUNY Oswego in 2019, so I learned it through Coursera, Udacity, papers, and open-source code alongside the degree. I still work that way: read a new paper, implement it, and adjust from what the results show. Training in the gym runs on the same loop: observe, adjust, iterate.

Basketball

I started playing in high school, joined a school tournament, and never stopped. It is also how I make friends every time I move: teammates at SUNY Oswego, pickup games by the lake in Gjøvik, and games with colleagues from DXOMARK in Paris. I still play today.

Kobe Bryant and Stephen Curry taught me that work ethic and adaptability matter off the court as much as on it.

By Lake Mjøsa, Gjøvik With teammates at SUNY Oswego Pickup game in Gjøvik With DXOMARK colleagues, Paris Watching the France vs Monaco final in Paris

Cooking

Mostly desserts. Cheesecake is my go-to; I also bake carrot cake and make pizza from scratch, dough included.

Cheesecake Carrot cake Homemade pizza

Photography, hiking & fitness

I take photos wherever I go. Studying across Norway, France, and Finland meant a lot of new places, from the northern lights over Gjøvik to Christmas markets in Alsace. My first dream destination was Yosemite; Europe came next.

I also hike and train at least four times a week, as much for discipline as for the body.

See the gallery →

Khang hiking on a mountain trail Northern lights, Gjøvik Hungarian Parliament, Budapest

I completed my master's in September 2026 and I'm based in Paris, open to Machine Learning and Research Engineer roles. The fastest way to reach me is email.