Multimodal learning
Inverse-LLaVA
Reverse the usual alignment direction: map text into a pretrained vision space.
45.6% fewer task-training examples than LLaVA-1.5’s two-stage recipe; excludes backbone pretraining.
Read case study
From research to working AI.
I work across multimodal learning, recommendation, and ML infrastructure, connecting new ideas with systems people can use.
Four projects. One journey from idea to impact.
Selected work
The work, unpacked.
Look inside the work: how models perceive, how systems adapt, and what it takes to make them useful.
Browse all projects Recommendation systems
Treverse Recsys
A recommender built for cold-start items and inventory that never stands still.
Learning → serving Time-safe graph learning, verified releases, and controlled experiments.
Read case study
Language & human behavior
AI for Negotiation
An LLM pipeline turns negotiation transcripts into structured evidence for behavioral research.
$5,000 → $3 estimated annotation cost per transcript in the reported study.
Read case study
ML infrastructure
FORGE
Shared delivery, monitoring, and rollback, with a runtime suited to each workload.
Cloud to edge Infrastructure, artifacts, and deployment state stay independently reviewable.
Read case study
02 / How I work
Research and production,
in one loop.
A useful model is only part of the story. I follow the question through evaluation, deployment, and what the system teaches us next.
Method / 05
Each stage opens the project that best demonstrates it.
01 / 05
Open project
Frame the question.
Ancient Mortars structures roughly 10 million particle images for comparison.
02 / 05
Open project
Fit the representation.
Inverse-LLaVA maps text into visual space inside the language model.
03 / 05
Open project
Make the comparison legible.
Negotiation transcript coding reports a cost drop from more than $5,000 to $3.
04 / 05
Open project
Publish a reviewable release.
FORGE defines six contracts across distinct ML runtimes.
05 / 05
Open project
Return field evidence.
Treverse joins controlled exposure with reproducible learning and serving.
03 / Experience
Curiosity,
put to work.
I’m Xuhui, an applied scientist at Treverse. My path connects academic research with the everyday challenges of building AI products.