Hi, I'm Truong Giang. I build machine learning that actually ships.
A London-based AI & ML consultant. Five years building machine learning across manufacturing, retail and healthcare — computer vision, RAG assistants, multi-agent systems — and I write about all of it as I go, including the parts that didn't work.
MSc Data Science (Distinction) · 5 yrs building ML · Good Clinical Practice certified · London, UK
Ways I can help
End-to-end, from a half-formed idea to a system your team can actually run. Pick a slice or the whole thing.
LLM fine-tuning & RAG systems
Custom LLMs and retrieval-augmented generation with grounded, low-latency answers. Multi-agent setups, enterprise chatbots, document processing, knowledge management.
Computer vision development
Image and video analytics for manufacturing and retail. Defect detection at 98% precision in 0.2s, object detection, visual inspection, and quality-control automation.
MLOps & pipelines
Production ML infra on GCP and AWS. Automated training with Vertex AI and Airflow, CI/CD, deployment, monitoring, and continuous retraining that doesn't fall over.
Predictive analytics & recsys
Forecasting, segmentation, and recommendation engines — turned into decisions and measurable growth, not dashboards nobody opens.
AI strategy & consulting
Roadmaps tied to business goals. Opportunity assessment, feasibility, and implementation planning — ROI-focused, with clear milestones and honest trade-offs.
Healthcare & medical AI
Clinical AI with the compliance to match. GCP-certified, with research on Parkinson's treatment prediction — medical data analysis and decision support.
Things I've built
A few projects across manufacturing, retail, and healthcare — plus what I build for fun.
Multi-agent customer support
A working multi-agent support system on Gemini — a Policy Agent (RAG) and a tool-using Ticket Agent — built end to end, then rebuilt in LangGraph to compare.
- ›Master + Policy (RAG) + Ticket agents
- ›15 scenarios, LLM-as-a-judge eval
- ›Open source, 8-part write-up
Fashion item detection
Led 3 engineers building the RetinaNet vision layer under outfit recommendation — accurate on its categories, archived when the data didn't transfer market.
- ›92% average precision, 19 categories
- ›Led a cross-functional team of 3
- ›Archived — a dataset-transfer lesson
Semiconductor defect detection
A two-stage defect-detection system for semiconductor QC with a full label-train-evaluate-infer desktop app — finished and ready to ship when COVID cancelled the client.
- ›98% precision in defect identification
- ›0.2s inference per image
- ›Never deployed — client cancelled
Parkinson's treatment research
MSc dissertation research on whether smartphone and wearable cues help treat Parkinson's-related drooling — 70 participants, four hypotheses, one salvaged dataset.
- ›70 participants across trial phases
- ›65% F1 on cue-timing recommendation
- ›Graded 90/100 (Distinction)
Vietnamese recipe chatbot
A RAG assistant over 800 Vietnamese recipes that reads a customer's basket to suggest dishes — built before RAG had a name or any tooling.
- ›Semantic retrieval over 800 recipes
- ›Vietnamese embeddings, done early
- ›R&D — no engagement metrics
Competitor price mapping
Multimodal SigLIP matching that mapped 6,000 SKUs across six competitor retailers daily — in production for over two years, feeding real pricing decisions.
- ›6,000 SKUs across 6 retailers
- ›2+ years in production, daily runs
- ›Led a team of 3 engineers
A bit about me
I'm Truong Giang, the person behind twentytwotensors. I started it because I like the whole arc — sitting with a messy problem, finding the model that fits, and seeing it run in production where it has to keep working.
My work spans NLP, computer vision, and cloud-native ML across manufacturing, retail, and healthcare. A price-mapping system matching 6,000 SKUs a day that ran in production for over two years; a semiconductor defect detector at 98% precision in 0.2s an image; a garment detection model at 92% average precision, where I led a team of three.
I bring the full loop — data engineering, model development, deployment, and the tuning afterwards — on GCP and AWS. And I write about it openly: the wins, the middling eval scores, and the projects that got archived. You'll find both kinds in the case studies below.
A price-mapping system that ran daily for 2+ years, and a BigQuery analytics platform teams used every day.
London-based, Good Clinical Practice certified, and used to handling sensitive health data carefully.
From data engineering to monitoring — the whole ML lifecycle.
An active blog documenting real projects, honestly.
From the blog
Notes from real projects — what I'm building, what worked, and the results I'm not going to dress up.
What it actually costs to self-host a coding model
I wanted to run my own coding model to save on a subscription. Priced across Qwen3.5, GLM 4.7 and DeepSeek, the honest answer surprised me: at modest usage self-hosting isn't cheaper than an API — it's the unlimited, private option. Where the crossover is, and the tok/s catch the spec sheet hid.
The score is about connections, not cells
Here is a way to fail at cell tracking: find every single cell, perfectly, and still score close to zero. Nobody asked where the cells are — they asked which cell became which. Part 1 of the Biohub series.
Build a scorer you trust before you build a model
You get five submissions a day. That is not an iteration loop, it is a rationing system. So I rebuilt the competition metric locally — and it predicted the leaderboard to within 0.008. Part 2 of the Biohub series.
Finding cells is the easy half — and the biggest win wasn't mine
A hand-written blob detector beat two well-known deep segmentation models — on accuracy and on runtime. Then I swapped in someone else's U-Net and gained more than everything I'd done combined. Part 3 of the Biohub series.
Greedy is a trap: linking cells with an integer program
Take the best link first and you can block two better ones behind it — 1.45 against a possible 2.55. The fix is to stop choosing one at a time. Part 4 of the Biohub series.
Let's build something
Tell me what you're working on. I'll reply within a day, and the first conversation — scoping out whether ML even helps — is free.