Chapter I
Enlist
Singapore Armed Forces2018 — 2020
I build the plumbing enterprise AI runs on
Software engineer at NCS in Singapore. Retrieval pipelines, agent tooling, and the backend work that decides whether any of it survives contact with real users.
Open to conversations about production AI.
Places that taught me something
Six years from simulation coursework to production AI, most of it in Singapore.
One engineer, the whole path from prompt to production
Most AI projects don't fail at the model. They fail at retrieval quality, evaluation, and the boring integration work nobody scoped. That's the part I do.
⌁
Rebuild chunking + reranking for an enterprise assistantshipped◇
Ground a healthcare knowledge tool in cited source documentsshipped◆
Wire tool-calling agents to internal APIs, with approval gatesin progress▤
Regression suite so a prompt change can't quietly break answersin progress▦
Ship the interface the whole thing is judged byshippedAsk me to walk through any of these in detail…
↑
Backend first
Python and C# services that hold up under load, with the schemas and tests to prove it.
Retrieval you can measure
If I can't show you the citation and the eval score behind an answer, it isn't finished.
Humans keep the keys
Agents draft and propose. Anything consequential waits for a person to approve it.
How I got here, in four chapters
Read them in order — each one is the reason the next made sense.
Chapter II
Learn
SIT · DigiPen2020 — 2024
Chapter III
Build
Accenture · NCS2023 — now
Chapter IV
Scale
Georgia Tech · AI Singapore2025 — 2026
Three things I shipped and what they cost
Client details stay vague on purpose. The engineering doesn't.
⌁
An enterprise assistant people actually trusted
A retrieval-augmented assistant over years of internal documents. The first version answered fluently and wrongly. Fixing that meant better chunking, a reranking pass, and refusing to answer when retrieval came back thin.
- Role
- Backend and retrieval engineering
- System
- Chunking → retrieval → reranking → grounded answer
- Guardrail
- Citations and abstention when retrieval comes back thin
- Stack
- Python · LangChain · Azure OpenAI
What's our retention policy for archived project files?
Archived project files are held for seven years from
project close, then reviewed before disposal. Two policies apply — the second one
supersedes the 2019 version.
records-policy-2023 · §4.2it-retention · §7
And for contractor files?
I don't have a source that covers contractors specifically.
I'd rather say so than guess — the closest document is §4.2 above.
◇
Evaluation that turns opinion into a number
Every prompt change used to be a debate. Now it's a run. A graded question set, retrieval hit-rate, and citation accuracy, checked on every merge — so a change that helps one answer and breaks four never reaches anyone.
- Role
- Evaluation tooling and API integration
- System
- 240 graded questions → CI run → regression gate
- Signals
- Citation accuracy, hit-rate, unsupported answers, latency
- Stack
- FastAPI · pytest · Azure Pipelines
Answers with a correct citation
0%
Graded questions
240
Retrieval hit-rate
+1488%
Answered without a source
−92%
Median latency
1.8 s
Regressions blocked
6
▦
Before the AI work, real-time systems
My degree was in real-time interactive simulation — C++, frame budgets, and code that has 16 milliseconds to be correct. Two years as a teaching assistant at DigiPen made me explain all of it out loud, which is still the fastest way I know to find what I don't understand.
- Role
- Simulation engineering and teaching
- Constraint
- 16.6 ms frame budget at 60 fps
- Transfer
- Performance budgets, reproducible debugging, clear explanations
- Stack
- C++ · C# · .NET · React
SimulationPhysicsRenderPresent
frame 1frame 60
Budget per frame
16.6 ms
Worst frame
−3.114.2 ms
Things I reach for
The working stack behind the projects, grouped by where it earns its keep.
Backend
Python, FastAPI, C#, .NET, REST APIs, schemas, and service-level testing.
Retrieval & agents
LangChain, Azure OpenAI, RAG, chunking, reranking, tool calling, and approval gates.
Quality & delivery
pytest, graded eval sets, citation checks, regression gates, and Azure Pipelines.
Systems & interface
C++, React, TypeScript, real-time simulation, profiling, and performance budgets.
The wider toolbox