AI and machine learning, built around your problem.
I'm Truong Giang, an independent AI and ML consultant based in London. I build retrieval assistants, computer-vision systems and data pipelines, drawing on work across retail, manufacturing and healthcare research.
London, UK · MSc Data Science (Distinction)
Ways I can help
Help with a specific part of your project, or with taking it from data and an initial prototype to a working system.
AI assistants & document workflows
Retrieval over your documents, assistants that use defined tools, and model adaptation where it fits the task. I work on the data, retrieval and evaluation as well as the conversation.
Computer vision & matching
Object detection, visual inspection and matching products across catalogues. I can help prepare the data, compare approaches and work through the constraints of the intended deployment.
Data pipelines & ML evaluation
SQL and Python pipelines, reproducible experiments, and tests for retrieval and agent behaviour. The aim is to make the results inspectable and the next engineering decision clearer.
Things I've built
Recent work on retrieval and agent evaluation, followed by earlier commercial projects and research.
Agentic evaluation framework (LLM-as-a-Judge)
An evaluation harness for a clinical-guideline research assistant, combining RAGAS, domain metrics, and checks on routing and identifier redaction.
- ›Retrieval, answer and workflow evaluation
- ›Investigated scoring and routing failures
- ›Research prototype, with limitations documented
Multi-agent customer support
A support-agent project for fictional UKConnect, with policy retrieval and booking tools. Built on Google ADK, then rebuilt in LangGraph to compare the approaches.
- ›Master + Policy (RAG) + Ticket agents
- ›15 scenarios, LLM-as-a-judge eval
- ›Open source, 8-part write-up
Earlier projects and research · 5 projects
Fashion item detection
Led a team of three building a RetinaNet model to detect garment categories for outfit recommendation. The project was archived after the dataset did not transfer to the target market.
- ›92% average precision, 19 categories
- ›Led a cross-functional team of 3
- ›Archived — a dataset-transfer lesson
Semiconductor defect detection
A semiconductor inspection system with a desktop app for labelling, training, evaluation and inference. It was completed but not deployed after the client project was cancelled during COVID.
- ›98% precision in defect identification
- ›0.2s inference per image
- ›Never deployed — client cancelled
Parkinson's treatment research
MSc research into smartphone and wearable cues for Parkinson's-related drooling, using data from 70 participants to investigate four hypotheses.
- ›70 participants across trial phases
- ›65% F1 on cue-timing recommendation
- ›Graded 90/100 (Distinction)
Vietnamese recipe chatbot
A research prototype that retrieves from 800 Vietnamese recipes and uses the contents of a customer's basket to suggest dishes.
- ›Semantic retrieval over 800 recipes
- ›Vietnamese sentence embeddings
- ›R&D — no engagement metrics
Competitor price mapping
An earlier retail project using multimodal matching to map 6,000 SKUs across six competitor retailers. It ran daily in production for over two years to support 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. My recent work focuses on retrieval assistants, agent evaluation and a small on-device audio detector. I write up the implementation choices and experiments alongside the code.
Before consulting, I worked on retail data pipelines and computer-vision R&D, including systems used in production and prototypes that did not reach deployment. The project pages explain those different outcomes.
My MSc research at Northumbria used smartphone and wearable data in a Parkinson's study. I have also completed NIHR Good Clinical Practice training. That background informs how I approach study data, evaluation and the limits of a model's results.
From the blog
Project notes on experiments, implementation choices and the results so far.
Why my AI agent was billing my Claude subscription as 'extra usage'
The clients used the same token, but appeared to use different balances. Matching the CLI's request fields stopped the extra-usage drain in my setup, without revealing the provider's internal billing rules.
Running three detectors for real, and stopping early
The chip model outperformed my DSP baseline in the live comparison. I also report the early stop, the incomplete evidence for the full exit rule and the limits of evaluating the retrained model.
Shrinking a CNN to 9.5 KB and putting it on a chip
A small CNN fitted the device where the cloud candidates had exceeded their budget. Getting it ready also meant checking the audio frontend, rounding differences and which days each evaluation model had seen.
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.
Project work from £300/day. Rates depend on scope and engagement length. Fixed-price quotes are available for clearly defined work.