hi, i'm junjie (jj).
a little about me
I build things that (hopefully) work, then keep tinkering until they feel right. I'm at Tencent in Singapore now. Before that, I shipped AI products at Didero in NYC and worked on multilingual AI research at AI Singapore.
This is my grasshut: a little patch of the internet for things I've made, lessons I've learned, and whatever grows next. Come on in.
work so far
- 01
Tencent
Software Engineer InternAug 2026 - nowI build internal tools that help product managers turn frontend ideas into working product changes.
- Built tools and agent skills that let 8+ product managers create, change, and validate production frontend code, cutting average feature turnaround from two days to a few hours.
- Worked with designers and product managers to bring 30+ pages onto a reusable design system.
- Cut a key page's render time by about 90% and restored reliable preview environments so the team could test changes with confidence.
- 02
Didero
Member of Technical StaffAug 2025 - Jul 2026Didero builds AI procurement software. I helped turn its agents into products that enterprise purchasing and accounts payable teams could use every day.
- Shipped controls that let customers guide AI agents and accept, edit, or reject proposed emails and ERP actions. Task acceptance rose from 40% to 80% as these controls rolled out across 20+ clients.
- Co-led an accounts payable product that matches invoices to orders, finds the right people to resolve discrepancies, and follows up by email; piloted it with 3 clients.
- Worked directly with 7 enterprise clients on ERP integrations and tailored workflows, adapting the product to how their teams actually worked.
- Built a trace explorer and improved caching and email intake so the team could understand agent behavior and run it more reliably at scale.
- 03
AI Singapore
AI Engineer InternMay - Aug 2025I worked on post-training for SEA-LION, a multilingual model built for Southeast Asian languages.
- Built a multi-agent reward pipeline to score and filter training data across 10 languages; the resulting model family ranked #4 of 55 on SEA-HELM.
- Improved code-judge accuracy from 50% to 80% while cutting scoring time from two minutes to under ten seconds.
- Compared reinforcement learning methods and connected the reward system to SEA-LION's training loop.
- 04
SUSS
Software Engineer InternJan - May 2025I explored practical AI tools for classroom teaching and information retrieval.
- Built retrieval systems and scrapers that could gather useful information from changing websites.
- Evaluated AI tools for classroom teaching.