cs @ ntu · singapore
$cat ~/about.txt
hi, i'm dharshan. i am an ai/ml and software engineer studying computer science at ntu.
I build things that people use. At Cyber Sierra, I was the sole developer of an AI agent used daily by the product team, from the full stack to deployment.
I check where AI gets things wrong. Built evaluations across four compliance frameworks and checks for unsupported product claims in Shopee Live Producer.
I make existing workflows better. Helped cut AI cost per trace by 45% at Cyber Sierra. At Netvirta, my tool cut size-chart processing from 4–5 hours to under 30 minutes per brand.
I dig into how things work. Fixed cache-token reporting in LiteLLM with regression tests. Merged upstream. I also explain my engineering decisions in writing, including how I built a voice shopping agent.
I take ownership and work with others. Owned Cyber Sierra’s agent end to end and led seven backend engineers building a platform for 900+ incoming NTU students.
Aug 2026 – Present
LLM Agent Development Intern
Supported by the CyberSG R&D Programme Office (CRPO) Internship Programme.
Built an eval harness in LangGraph and Langfuse across four compliance frameworks, plus a paired benchmark of a baseline against the DSH agent. I also deployed tool-history replay to Cyber Sierra's codebase knowledge agent, which ran 20.1% cheaper and 25.5% faster on 6-turn Langfuse traces.
May – Aug 2026
AI Engineer Intern
Sole developer on a codebase knowledge agent, full stack and deployed on AKS. 2 to 3 people on the product team use it daily. I also helped cut average cost per trace by 45% on an agent platform.
Dec 2025 – Apr 2026
Backend Engineer Intern
Moved 6 scrapers from TypeScript to Python on AWS SQS, part of the Python infrastructure behind 10,000+ daily scraping jobs.
Feb – Nov 2025
Backend Engineer Intern
Built an LLM extraction tool for size charts that cut manual processing from 4 to 5 hours to under 30 minutes per brand, across 80 brands.
Jan – Jun 2026
Head of Technology, Backend
Led 7 backend engineers building the platform for 900+ incoming students.
[in progress]
AI Safety Researcher in a research mentorship program. A mechanistic-interpretability study of why chat safety judges fail: I built the PyTorch inference pipeline and batch screening, and use activation patching.
[in progress]
Undergraduate Researcher with Prof. Wang Chen, studying cost, latency, and quality in multi-agent workflows.
A short-video recommender on the public KuaiRand-Pure dataset: two-tower retrieval with FAISS, a caption safety filter, then LR, DeepFM and MMoE rankers behind a FastAPI service. Over 5 seeds, Recall@200 beat popularity by 14% and ItemCF by 25%, with paired bootstrap CIs that exclude zero. A load test ran 2,890 requests with 0 failures at 100 ms p99. It taught me that a better model score is not the same as a better system.
A toxicity classifier from a QLoRA fine-tune of Qwen3-4B. On a frozen test set it reached 0.62 F1, against 0.27 for the base model and 0.35 for TF-IDF logistic regression. LoRA weights are on Hugging Face.
3rd of 40 teams at the Sea × OpenAI hackathon. An AI co-producer for live commerce, built because customer questions kept going unanswered during selling streams. I owned the agent orchestration, including checking claims against available data before a response goes out.
Merged PR #31076. Anthropic-compatible cache-token reporting across streaming paths.