Qihan BaoAI Research Engineer

Between models and the real world

I work on LLM inference and evaluation, and build applications around models, agents, and backend systems.Stay curious. Keep thinking.

See my work Get in touch

Conceptual particle trajectories converging into a layered system — Illustration; not experimental data.

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?

Nutri-flow began as my capstone project. I started with FoodSeg103 semantic segmentation, then built the asynchronous analysis pipeline, agent workflow, web and mobile apps, and deployment around it.

The prototype runs end to end, combining segmentation, nutrition retrieval, and user memory. Its estimates are rough and are not medical advice.

What I worked on

I built the model, agent workflow, backend services, client applications, and deployment. The model uses a pretrained Swin Transformer with learnable multiscale fusion; RabbitMQ queues the analysis jobs, and LangGraph manages the stateful workflow.

Illustration; not experimental data.

I've focused these descriptions on the parts I actually worked on. The linked repositories and papers have the implementation details and results.

Jet pT reconstruction in simulated high-background events

Jet pT reconstruction

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

The study embedded PYTHIA6 hard-jet events in a simulated thermal background, then clustered the resulting jets with FastJet anti-kT.

We compared an MLP, random forest, and linear regression with area-based subtraction. The results here come from Monte Carlo simulation, not collision data.

What I worked on

I led the four-person project and was a co-first author. 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?

I trained LoRA adapters for nine portrait styles on a pretrained diffusion model and ran the image-to-image experiments.

The paper also includes conditional diffusion and F-ESRGAN enhancement, which were part of the broader team project.

What I worked on

I was a co-first author in the three-person team. My work focused on LoRA training, portrait-generation experiments, and analysis of the results.

Technologies used: DiffusionLoRAimage-to-imageF-ESRGAN

Read the paper

What I look for in an experiment

I want the question to be clear, the evaluation to be easy to check, and the claim to match the evidence. If the evidence is not strong enough, I would rather stop there than force a conclusion.

Mar 2026 to Apr 2026

A test-time compute experiment that did not become a paper

In spring 2026, I worked with a PhD student at a German university on a small set of Qwen2.5 sampling experiments on MATH and AIME 2024. I used Slurm to run the cluster jobs and compare different inference budgets.

The sample was too small for a stable conclusion, so we stopped instead of trying to turn it into a paper.

The main lesson was to get the evaluation setup and experiment records right before spending more compute.

Education and experience

My degree, an industry training project, and the next step I'm still arranging.

Sep 2022 to Jul 2026

Southwest University

Bachelor of Engineering in Software Engineering

Graduated in July 2026.

Mar 2025 to May 2025

Chinasoft Excellence, Chongqing

University-industry capstone program

I was the project manager and agent design lead in a six-person team. I coordinated delivery and designed a MaxKB smart-home agent/RAG workflow with tool and data-source routing.

Enrollment date TBD

University of California San Diego

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

I've been admitted to the ECE master's program and am still arranging my enrollment date.

What I'm working on now

Most of my time is going into LLM inference and evaluation. I'm also maintaining older projects and building a diet-tracking mini program for my family.

LLM inference and evaluation

I'm exploring test-time compute, sampling strategies, and evaluation design, with an emphasis on keeping clear experiment records.

Putting models into applications

I'm working on the engineering around complete applications: how to split services, queue work, and fall back when something fails.

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 small interactive piece about orbit, energy, and choice.

I designed and built it myself, with AI help on parts of the implementation.

Enter Starfall Circuit

If you'd like to talk about any of this, get in touch

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

Send an email Visit GitHub