Jiajun Lim
AI engineer, full-stack: frontend, backend, and the infrastructure underneath.
I picked AI as my major before ChatGPT made it the obvious choice, mostly on a hunch that it would end up mattering. I still think it's the biggest lever we have for taking tedious work off people's plates, and that's the thread running through everything I build. Most recently: a multi-agent chatbot that answers customer questions for an investment platform and a digital bank from one core engine.
Stella, a multi-agent chatbot for financial services
Stella answers customer questions for an investment platform and a digital bank from one core engine, orchestrated with LangGraph. It handles the questions that come up again and again, taking real volume off the customer service team's plate so they can focus on what actually needs a person. Answers are automatically checked and can revise themselves before reaching the customer, with a graceful handoff to a human when they can't, and personal data is detected and redacted before anything is processed. It searches an internal knowledge base on Azure AI Search and Google Vertex AI, falls back to live web search with Gemini grounding when that comes up short, streams responses back in real time, and traces every run end-to-end in Langfuse.
→ Also diagnosed and fixed a production incident where heavy concurrent load could stall chat responses for over a minute. I reworked the streaming pipeline so it holds steady under the same load, with responses starting almost immediately.
Fintech lending integration platform
On a client engagement integrating a lending partner into a fintech platform, I owned the backend, a FastAPI service handling the partner's callback infrastructure for loan and disbursement notifications, authenticated with bearer tokens and IP allowlisting. I built the SFTP integration that exchanges PGP-encrypted payloads with the partner, and solo-provisioned and ran the infrastructure: a DigitalOcean VM behind Nginx, an egress proxy for outbound IP control, and Cloudflare tunnels fronting UAT and prod.
CVM stage classifier
Before the agent work, I built a deep learning classifier for cervical vertebral maturation staging from dental X-rays, a diagnostic task orthodontists use to time treatment. I fine-tuned a ResNet backbone (frozen first, then progressively unfrozen), used L1/L2 regularization to keep it from overfitting on a modest dataset, and added Grad-CAM heatmaps and confidence scores through a Streamlit interface, so a clinician sees which part of the X-ray drove each prediction instead of trusting a black-box number.
Comfortable across cloud and bare infra
Azure and GCP for the managed side, Kubernetes and DigitalOcean VMs for the bare-metal side. I build across CLIs, Claude Code, Codex, Grok CLI, Meta's Muse Code, and whatever else is worth trying, switching often and having one model judge another's work before it ships. I like closing the agentic loop myself, automating the path from code to shipped, CI/CD included.
Outside of work, I follow macro and policy, and I like puzzling through options and equities. I'm drawn to decentralization for a similar reason: value should move as easily for someone without banking infrastructure as for anyone else. Years of K-dramas and K-pop got me to basic conversational Korean. I've recently picked up video editing, and filter coffee is a daily ritual.