backend engineer · singapore
I'm a backend engineer working on onboarding systems — and, lately, exploring the AI tooling around them on my own time: RAG pipelines in production, and an internal memory layer for AI coding agents that some of my team has started using too.
I try to keep things simple and avoid overengineering. AI is something I'm still figuring out as I go — I just enjoy poking at it when I run into a real problem worth solving.
skills-for-microservices Claude Code skills
Open-source skills that help AI agents work across microservice architectures: tracing a frontend API call to a scoped backend change plan across services, generating per-service architecture specs (with domain-pollution and dead-code detection) that agents can consume as context, and syncing a fleet of repos at once.
Surviving a 52x Traffic Spike on an LLM Pipeline jul 2026
Incident notes: the model held up — the part that broke was the human fallback queue, the one piece of the system that doesn't autoscale. On timeouts as policy decisions, and draining a backlog without hurting live traffic.
Designing an LLM Routing System jul 2026
Deep-dive design notes: who decides which model serves a request, why cost is computed in four places but only one is real, and why routing is per-conversation once prompt caching enters the picture.
Teaching AI Agents to Work Across Microservices jun 2026
Why cross-service work is where coding agents struggle most, and the three skills I open-sourced to help — tracing API calls across service boundaries, and documenting which flows are legacy so agents stop building on the wrong ones.
What I Learned Building a Memory Layer for AI Coding Agents may 2026
Lessons from building an internal tool nobody asked for, that most of my team now uses daily — on starting simple, earning adoption, and why there's no one true AI workflow.
How AI Coding Agents Actually Work feb 2026
The ReAct loop in plain language — thought, action, observation, with an animated step-through — and how an agent figures out which tool to reach for.
Experimenting with LLM-assisted engineering workflows — my working conclusion is that the right workflow depends on the engineer's seniority, familiarity with the codebase, and the type of ticket, not on any single tool. Currently building a Graphify-inspired codebase knowledge graph that plugs into tooling Graphify doesn't support, exposed over MCP as a shared team resource — and extending my memory tooling with semantic retrieval.
From university and internship days — less relevant to what I do now, kept for the record: GraphVisualisation (C++ force-directed graph drawing, Barnes–Hut optimised — cut genome graph rendering latency 65% during my A*STAR internship) · dbms (a simplified DBMS with B+ tree indexing, from scratch) · SequenceAligner (a simplified minimap2 implementation) · Dynamic Graph Clustering (NTU final year project: GCN-based node embedding and clustering).