Work Experience
Where I've worked, and what I built there.
Garuda Robotics
Case study
AI Engineer Intern · Singapore
Building an AI assistant that plans drone inspection flights. The operations team describes a job in plain words, answers three questions, and gets a flight plan saved in the planning platform, already checked against the area they are cleared to fly. A person reviews it and a licensed pilot flies it. Before this, someone entered around twenty flight settings by hand. It has been running in production since October and was demoed to visiting executives on 23 October.
I built the tool service the assistant works through, and the desktop app the team uses. The service gives the model 12 tools, is covered by 298 automated tests, and enforces the limits on what any request can do. I also chose and configured the security gateway in front of it, after comparing two open-source options.
Most of the work has gone into the safety side. The thing at the other end of this is a drone, not a document, so a flight plan outside someone's approved area has to be impossible to create, not just unlikely. That's written up as a case study.
I also connected the assistant to live drone video, so it can raise fire and perimeter alerts and write short situation reports, and moved the service onto Kubernetes on the company's own servers, with Docker and an automated test-and-deploy pipeline. That work is the groundwork for opening the tools up to outside developers.
- Agentic AI
- MCP
- TypeScript
- Electron
- Local LLMs
- Kubernetes
SP Digital
Case study
Data Science Intern · Singapore
Six months on the guardrails for an internal assistant used by frontline staff at a utility, plus the adversarial test suite that showed whether they worked. Unsafe responses dropped roughly 60%.
The people using it are out doing physical work on the grid. Some of them can see things the others shouldn't, so a lot of the job was working out what the assistant says to whom.
This was the first thing I'd built where someone was paid to break it, and it changed what I think the work is. I've written that up properly as a case study.
- LLM Systems
- Guardrails
- Enterprise AI
- Evaluation
Lecture AI
Case study
Co-Founder · Singapore
Co-founded a startup that turned lecture recordings into bilingual study notes. We built what students asked for and then hit a problem that had nothing to do with the product.
- AI
- NLP
- RAG
- Startup
- Product
KPMG
AI Labs Intern · Gurugram
Built an assistant that let a twelve-person consulting team ask questions of their 200+ internal documents instead of searching for them, with every answer linked back to the documents it came from. Lookup time dropped by 40 to 50 percent.
A twelve-person consulting team kept everything it knew in a shared document store. Over 200 files, findable only if you already knew roughly what you were looking for and what someone had called it. The cost wasn't dramatic. A few minutes lost, several times a day, across everyone, plus the invisible cost of redoing work because you couldn't find the thing that already existed.
I built a system that let them ask a question instead of searching. It's an agentic setup rather than a single retrieval pass. It decides what to look for, retrieves, and goes back for more if the first pass doesn't answer the question. That mattered here, because consulting questions usually need pieces from several documents rather than one good paragraph.
The chunking took a few attempts to get right. Long PDFs split naively produce chunks that end mid-argument, and the model fills the gap with something plausible. I moved to 1000-token chunks with 200 tokens of overlap, so a point that straddles a boundary appears whole in at least one chunk. That fixed most of the wrong answers I was seeing.
The bigger thing I underestimated was trust. Consultants didn't just want the answer, they wanted to know where it came from, which is fair when their name goes on what they send a client. So every response returns the top four supporting documents as links alongside it, and spreadsheets are handled properly so numbers come back as numbers rather than mangled text.
Retrieval made the answers fast. Being able to check them is what made anyone use them. Lookup time dropped by 40 to 50 percent.
- RAG
- LangChain
- Azure OpenAI
- Enterprise
AlygnAI
SDE Intern · Remote (US)
A remote internship with a Bay Area startup, run alongside a full-time one in India. Same summer, two jobs.
I ran this alongside KPMG. Same summer, one in person and full time, this one remote. I don't recommend it, but I wanted it badly enough to try.
The reason was specific. I'd read a lot about how Bay Area startups work and I wanted to find out whether the version in my head matched the real thing. So I spent that spring cold-applying to American startups from the other side of the world, which is a slower and more discouraging process than it sounds, and eventually got one.
What I found was mostly what I'd hoped for. The team was small and moved quickly, and nobody had time to hand me a scoped task and check on it. I ended up owning a migration from a prototype built on a no-code tool to a real backend, start to finish. That's more ownership than I'd have got anywhere with more people in it, and it's the thing I'd point at from that summer. I also compared two ways of adapting a language model to their domain, fine-tuning it on their data against retrieving from their documents at query time, and wrote up the tradeoff for the founding team's architecture decision.
The double summer was probably one summer too many. It was also the only way I was going to find out what I wanted to know.
- FastAPI
- Auth
- LLMs
- Startup
StatusNeo
SWE Intern · Gurugram
Built REST APIs in Spring Boot for an enterprise banking client, with token-based login (JWT) and role-based access control, and wrote the JUnit tests for them.
- Java
- Spring Boot
- REST APIs
- Enterprise