Now

2026

Jan

• Played with OpenCode.
• Reading Recursion via Pascal
• Reading Make Time.
• Brought many old posts back to life.
• Improved the sta­bil­ity of my LLM-based in­ges­tion pipeline in­clud­ing bet­ter mes­sage queue con­fig­u­ra­tion, im­ple­mented retry mech­a­nism, and fall­back logic.
• In the process of mov­ing to Linux, Windows now be­comes un­bear­able for daily use.
• Developed a new prod­uct un­der­stand­ing model for gen­eral prod­uct. It’s a multi-step agent that lever­ages pre­de­fined tax­on­omy to ex­tract prod­uct at­trib­utes, gen­er­ate prod­uct de­scrip­tion, and cat­e­go­rize prod­ucts.

2025

Dec

• Completed my study at NUS. What a jour­ney it has been!
• My pro­posed LLM-enhanced in­ges­tion pipeline has gone live. Some un­ex­pected is­sues popped up dur­ing ini­tial launch.
• Developing LLM agents for gen­eral prod­uct in­ges­tion pipeline. Thought about us­ing knowl­edge graph to en­hance the model ca­pac­ity, but I did not have enough time to ex­plore this di­rec­tion.
• Played with Raku.
• Reading:
  • Introduction to Recursive Programming: my re­view.
  • Recursion via Pascal
  • How to Take Smart Notes

Nov

• Moved my old con­tent to the new place (which is this one). Still in progress though since I in­tend to move some an­cient posts from 201x era to here as well.
• Working on op­ti­miz­ing ma­trix mul­ti­pli­ca­tion, ap­plied sev­eral tech­niques such as block­ing, loop un­rolling, and SIMD in­struc­tions.
• Working on ques­tion clas­si­fi­ca­tion us­ing graph ma­chine learn­ing.
• Read:
  • Linear Algebra for Data Science, Machine Learning, and Signal Processing
  • System Performance: Enterprise and the Cloud, 2nd Edition

Oct

• Experimented with Quartz. I’m go­ing to use it for my up­com­ing “Fun with Data Structures & Algorithms” (Definitely need a bet­ter name) se­ries.
• Been work­ing on Graph Neural Networks for ques­tion clas­si­fi­ca­tion. Played with PyTorch Geometric quite a lot.
• Looking for bet­ter note-tak­ing apps. In the past, I used the com­bi­na­tion of UpNote, Obsidian, and RemNote. Lately, I found Logseq is a great al­ter­na­tive to the above apps.
• Found sphin­x_lit­er­ate which maybe the best lit­er­ate tool at the mo­ment. Look at this beauty.
• Read:
  • Linear Algebra for Data Science, Machine Learning, and Signal Processing
  • System Performance: Enterprise and the Cloud, 2nd Edition

Sep

• The new re­quire­ment from Google for apps to pub­lish to store, i.e, sup­port­ing 16kb page size took me sev­eral days to fig­ure out.
• During PR re­view, I need to re­vise the cur­rent C4 di­a­gram & se­quence di­a­grams of the new ser­vice. Previously I used mer­maid for draw­ing, but later found plant­text is much bet­ter.
• Studied spec­tral the­ory and its ap­pli­ca­tions in ML such as spec­tral clus­ter­ing, Graph SVM, and rec­om­men­da­tion.
• Studied queue­ing the­ory for com­puter sys­tem per­for­mance analy­sis.
• Read:
  • Linear Algebra for Data Science, Machine Learning, and Signal Processing
  • The Art of Computer Systems Performance Analysis
  • Queueing Systems, Volume 1: Theory

August

• Spent con­sid­er­able time on model de­ploy­ment. I wanted to break the ex­ist­ing de­sign of our ML in­fra­struc­ture which is out­dated and trou­ble­some to de­ploy. Went to MLServer, sim­pli­fied the de­ploy­ment by re­mov­ing the de­pen­den­cies on other ser­vice, also re­moved sev­eral re­dun­dant com­mu­ni­ca­tion be­tween ser­vices. At the end, I still need to up­date 4 dif­fer­ent ser­vices.
• Started the fi­nal se­mes­ter at NUS. I took 2 courses which seem un­re­lated to each other: Graph Machine Learning, Computer System Performance Analysis.
• Reading:
  • Test-driven de­vel­op­ment in Go.
  • Cloud Native Go.
  • Learning Go: about to fin­ish, good book so far. Too ver­bose for my taste.
  • tmux 3: Quick read, learned some new tricks.
  • Systems Performance: Enterprise and the Cloud.
  • Graph Neural Networks in Action.

July

• Designed and im­ple­mented an LLM-based prod­uct in­ges­tion ser­vice. Due to the la­tency of LLM in­fer­ence, in­te­grat­ing this ser­vice with the ex­ist­ing prod­uct in­dex­ing pipeline would vi­o­late cur­rent SLA re­quire­ments.
• The new ser­vice is im­ple­mented in Golang (previously I wanted to use Python - but app size may af­fect ser­vice startup time). Some books helped me a lot:
  • Test-driven de­vel­op­ment in Go.
  • Cloud Native Go.
  • Learning Go.

Jun

• On-device ML de­ploy­ment: Upgrade our cus­tomized Mediapipe for the lat­est Unity (6.1) as well as the new SDK from Android & iOS. The in-house ver­sion is way too old and has many is­sues, I will prob­a­bly need to fork me­di­apipe again and re-im­ple­ment our SDKs.

• Develop a prod­uct sum­ma­rizer ser­vice us­ing ChatGPT. POC is easy, de­vel­op­ing a pro­duc­tion-ready ser­vice is not. The main chal­lenge is the fact that LLM is not a “real-time” ser­vice, and to save cost, we may need to use Batch API. For that, a whole new ser­vice needs to be de­vel­oped, fo­cus­ing on batch pro­cess­ing and han­dling race con­di­tion be­tween data­base up­dates and batch pro­cess­ing.

• Read:

  • AI-Powered Search.
  • Algorithmic Thinking.
  • Distributed Services with Go

May

• Read:
  • AI-Powered Search.
  • Algorithmic Thinking.

Apr

• Finished both Cloud Computing & Text Mining courses. Both are prac­ti­cal and I re­ally liked them, al­though I wish I could have spent more time on some top­ics taught in lec­tures.
• Worked on:
  • Non-English prod­uct search: adding non-Eng­lish sup­ports to the in­dex­ing back­end and au­to­com­plete ser­vices.
  • Synonym set sup­port: It’s more about al­low­ing mer­chan­dis­ers to de­fine their own prod­uct tax­on­omy, the most tricky part is how to adapt vec­tor search to the cus­tomized de­f­i­n­i­tion from users w/​o re-train­ing mod­els for each cus­tomer.
• Read:
  • AI-Powered Search.
  • Algorithmic Thinking.

Mar

• Spent a con­sid­er­able ef­fort to hunt down search bad cases re­ported by cus­tomers. Reproducing is­sues is easy, but ac­tu­ally pin­ning the root cause is a com­pletely dif­fer­ent story. Long story short, it is due to wrong ex­pand­ing terms for sparse mod­els.
• Developed new code­base for em­bed­ding fine­tun­ing. Loved uv af­ter com­pletely switch­ing to it. I also re­ported some bugs to neural-cherche. Still strug­gled to get good per­for­mance on Splade, SparseEmbed gave me some hope, but not much.
• Reading
  • Introduction to Information Retrieval
  • Amazon Web Services in Action: Assisted in com­plet­ing as­sign­ments.
  • AI-Powered Search:  Engaging read, though some chap­ters were slightly dis­ap­point­ing.
• Traveled to Bangkok: What a jour­ney. Much fun, best food in SEA (aside from VN), and got scammed.
• Study:
  • Tried some event de­tec­tion mod­els and im­ple­mented event dedu­pli­ca­tion. I don’t like the idea of us­ing LLM for every­thing, but LLM ac­tu­ally is the best out-of-the-box so­lu­tion for event de­tec­tion.
  • Lots of cod­ing for the other pro­ject: learned more about OpenAPI, FastAPI, mon­goDB …

Feb

• A lot of cod­ing & hack­ing:

  • Developing back­end ser­vices for sport fa­cil­ity rec­om­men­da­tion.
  • Play around with event de­tec­tion and knowl­edge graph con­struc­tion.

• Reading

  • Speech and Language Processing
  • DNF Mastering Regular Expressions. My short re­view

Jan

• Working on . DNF:

  • Spell checker - I’ll write a post about mod­ern de­sign of spell checker for prod­uct search.
  • Introduce per­son­al­iza­tion to rerank.
  • Try LoRA on lan­guage mod­els.

• study@NUS: this se­mes­ter I’ll take Cloud Computing and Text Mining courses.

• Reading

  • Cloud Computing
  • Introduction to Information Retrieval
  • Programming Pearls: Will go through all ex­er­cises in the book.

• Done:

  • DevOps, DataOps, MLOps: 2/5 - NOT Recommended.
  • GNU Make Book: Review.

2024

Dec

• Reading The Silmarillion.
• Working on thresh­old op­ti­miza­tion.
  • Accuracy is main­tained while per­for­mance im­proves be­tween 50 – 80%, the first fea­ture is re­leased at the start of Y25.
• Improving multi-lin­gual search qual­ity by fine­tun­ing BGE-M3.
  • It was a huge pain 😠 😠. I stopped work­ing on it, the train­ing is not sta­ble with some ran­dom spikes of gra­di­ents that com­pletely de­stroys the train­ing.
• Studying: MLOps, AWS.
• Overhauled my home­page. I like the sim­plic­ity of the new theme.
  • Also added vale to my work­flow.
  • Decoupled the de­ploy­ment of lora from the home­page.
  • Upgraded to 11ty v3.
  • Added Github com­ments
• (VN) Reviewed Dune.

2023

May

• Playing around with SAM.
• Reading Designing Data-Intensive Applications. It is wayyyy over­due. I should have read this book long be­fore.
• Reading The Algorithm Design Manual. A fun read so far, most of the ex­er­cises can be prac­ticed on Leetcode.