KianAI Research Engineer

Between models and the real world

I study LLM inference and evaluation, and build applications with models, agents, and backend services.Stay curious. Keep thinking.

See my work Get in touch

Conceptual particle trajectories converging into a layered system

Selected work

From a meal photo to a working application

Nutri-flow

What can you build once a model knows what is on the plate?

My capstone connects meal photos, food segmentation, nutrition retrieval, and user history to everyday meal suggestions. The full workflow runs on web and mobile.

What I worked on

I built and deployed the full application. I trained a Swin Transformer segmentation model with learnable multiscale fusion on FoodSeg103, then connected the analysis pipeline with RabbitMQ and LangGraph.

Nutri-flow web app showing segmented food regions and categories — Food segmentation on the web, meal uploads on mobile. Open either image for the full screenshot.

Meal preview and upload screen in the Nutri-flow mobile app — Food segmentation on the web, meal uploads on mobile. Open either image for the full screenshot.

The projects I enjoy most start with a hard question and end with something I can test, use, or ship.

Machine learning in a simulated thermal background

Jet pT reconstruction

Can machine learning improve jet pT reconstruction when the background is noisy?

We embedded PYTHIA6 events in a simulated thermal background and clustered jets with FastJet anti-kT, then compared MLP, random forest, and linear regression against area-based subtraction.

What I worked on

As project lead and co-first author in a four-person team, I coordinated the study and contributed to the code, experimental design, figures, and paper.

Technologies used: PYTHIA6FastJetanti-kTMLPRandom Forest

Read the paper

LoRA portrait experiments in nine styles

PSDiffusion

Can LoRA produce different portrait styles without retraining the full diffusion model?

The paper combines conditional diffusion, LoRA style adaptation, and F-ESRGAN enhancement in a portrait-generation pipeline.

What I worked on

As co-first author in a three-person team, I trained LoRA adapters for nine portrait styles, ran image-to-image experiments, and organized the results.

Technologies used: DiffusionLoRAimage-to-imageF-ESRGAN

Read the paper

Questions worth testing

When one setting changes, why does the result move? I start with an evaluation plan, then record and compare each run.

Mar 2026 to Apr 2026

Qwen2.5: inference budgets and math reasoning

With a PhD student at a German university, I explored Qwen2.5 sampling on MATH and AIME 2024. I ran the cluster jobs with Slurm and compared inference budgets.

This was a small-sample exploration, without a stable conclusion yet.

Education

Sep 2022 to Jul 2026

Southwest University, China

Bachelor of Engineering in Software Engineering

Enrollment date TBD

University of California San Diego

Admitted to the M.S. program in Electrical and Computer Engineering

What I'm working on

More experiments, ongoing projects, and small tools for everyday life.

LLM inference and evaluation

Exploring test-time compute, sampling strategies, and evaluation design to compare how different settings perform.

Models in applications

Working through service boundaries, asynchronous jobs, and failure recovery to keep model-powered applications running reliably.

home-daily-diet

A diet-tracking WeChat Mini Program for my family. The core features are working; voice input and LLM-assisted recipe entry are still in development.

View on GitHub

Side project

Starfall Circuit

A six-minute space roguelite about collecting upgrades, surviving enemy waves, and reaching the final encounter.

Enter Starfall Circuit

Let's talk

I'm always happy to talk about LLM inference and evaluation, applied machine learning, agent systems, or systems engineering.

Send an email Visit GitHub