Liu Junxi · AI Product Manager

Turning AI capability into products people actually use.

Work

ByteDance CIS AI Full-Stack Project Challenge · Outstanding Project Award

RivalRadar

Outstanding Project Award, ByteDance CIS AI Full-Stack Project Challenge. Solo full-stack, core version in three weeks. Turns competitive research from a black-box long report into a reviewable decision process where every conclusion traces back to its evidence

Case study Live demo GitHub

3 wkscore version, built solo for the competition

4×specialized agents in one loop

402tests, all green on CI

88.4%Macro-F1 on evidence-support verdicts (post-contest eval)

Conclusion gating

Every comparison and decision recommendation binds a source and an exact quote. Citation numbers are generated and validated by the template, so the LLM never touches them. The model then grades support cell by cell in three states. Claims that cannot hold up get cut, and missing evidence stays visibly blank. An empty cell beats a forced answer.

QC as curator

The first QC could only reject the run as a whole and once misjudged a nearly complete matrix as insufficient data. Rebuilt with tiered disposition, it sends evidence gaps back to collection for targeted top-ups, structural issues back to analysis for rework, and cuts claims that cannot hold up. The trust state of every conclusion is shown to the user.

Replayable process

Queries, sources, per-cell checks, and self-correction loops stream to the workbench live. Refresh the page and the whole run replays. Trust in agent products comes from process transparency.

Rule gate → LLM verdict

Exact quote bound✓
Numbers template-generated✓
LLM never touches numbers✓
LLM supportFull

Unsupported claims get cut · missing evidence stays blank

Tiered disposition

Full supportPass
Partial supportMark boundary
UnsupportedCut

No death penalty for content. Show the trust state.

Live workbench

Fetch37 evidence items
Checkcell by cell
Correctself-correction loop
Audittrail fully written

Refresh the page, replay the whole run

2026-02 · Time-series Foundation Model · Independent Build · Streamlit

Crypto-Pilot

A zero-cost, GPU-free local crypto forecasting terminal built on a time-series foundation model, a complete independent exercise in foundation-model productization

Case study Live demo GitHub

0no API key, no GPU, no subscription

3×product calls: fully local, dynamic threshold, live controls

30dbacktest baseline shipped in the repo

AllBinance spot pairs, open to try

Experience

2026

Tencent Cloud Contact Center (TCCC) · Voice Agent Product Line

AI Product Manager Intern · 2026-06 – Present

Product work on voice AI agents at Tencent Cloud Contact Center (TCCC), spanning the agent quality evaluation system, the building-efficiency toolchain, and industry demo operations.

2025

Sichuan Center for Disease Control and Prevention

Algorithm Intern (“AI+” HIV Risk Assessment, University-Enterprise Joint Project) · 2025-07 – 2025-12

Helped design and ship an intelligent HIV regional-risk assessment system: F1 up 17.8% over the XGBoost baseline; a 90-district batch runs end-to-end in under 60 seconds (183 districts province-wide); piloted in 4 city/county CDCs.

I'm Liu Junxi, a software engineering grad student at UESTC. By day I work on voice AI agents at Tencent; off hours I turn ideas into running products. This site itself, content and code, is maintained by me and AI agents together.

You made it this far. Let's talk.

EMAIL2685949814@qq.com GITHUBgithub.com/Billkst

Class of 2027 · Currently interning at Tencent

billkst.