Skip to portfolio

AI engineering · Software engineering

Useful AI. Thoughtfully engineered.

I’m a Computer Science undergraduate at NTU. I build agentic tools, evaluate their reliability, and engineer the systems behind them.

Industry experience at Cyber Sierra, eLife and Netvirta. Agent evaluation at CRPO, NTU. A merged contribution to LiteLLM.

View selected projectsEmail Dharshan ↗Download resume ↓

Projects

The Hive

A few systems I’ve helped build, and the decisions behind them.

Agent orchestration3rd of 40 teams · Sea × OpenAI 2026

Shopee Live Producer

A backstage assistant that helps livestream hosts answer product questions without inventing product claims.

The problem
Busy livestream chats mix product questions, casual messages and risky claims. Hosts need help deciding what deserves a response.
My contribution
I owned multi-agent orchestration: routing buyer comments into four action classes and validating generated claims against database fact IDs.
Key decision
The model proposes an answer, escalation, warning or no action. A separate grounding check validates cited product facts before a reply is posted.
Outcome
Our team placed 3rd of 40 teams at Sea × OpenAI 2026. The repository includes the host/buyer flow and a reproducible demo script.
  • TypeScript
  • LangChain
  • DeepAgents
  • Supabase
Source & demo guide (opens a new tab)
Sources & scope

Contribution and placement: supplied resume. Architecture and demo instructions: project repository. No production adoption claim.

Conversational retrievalTikTok TechJam 2026 · Team TechBros

Beeline

A shopping copilot that remembers changing preferences and recommends products from a real catalogue.

The problem
A shopper changes their mind across a conversation. Retrieval needs to preserve the constraints that still matter and return catalogue-valid products.
My contribution
My documented slices covered the offline agent contract, live dense retrieval, hybrid fusion, fusion training, model benchmarking and reproduction. I also contributed the ranking tournament and robustness checks.
Key decision
Separate intent state from retrieval and ranking. Connected model stages have deadlines and fall back to a valid local result when a provider fails.
Outcome
The repository reports a 0.806 TechnicalScore on 200 public evaluator sessions, compared with 0.756 for its offline configuration, and publishes an executable evaluation path.
  • Python
  • Qdrant
  • Hybrid retrieval
  • OpenAI
View source (opens a new tab)Watch 3-minute demo (opens a new tab)Project submission (opens a new tab)
Sources & scope

Team-reported public benchmark, not independently rerun for this portfolio. Individual slices are attributed in the repository. No private-evaluation result is claimed.

Open source · ReliabilityMerged upstream · PR #31076

LiteLLM contribution

A fix for prompt-cache token usage that disappeared in the Anthropic Messages adapter.

The problem
Lost cache accounting made usage reports incomplete across streaming and non-streaming message responses.
My contribution
I unified usage conversion for both response paths and added regression tests for cache fields and delayed streaming usage.
Key decision
Keep one usage-conversion path and test the timing edge case directly, so both interfaces preserve the same accounting fields.
Outcome
The fix was merged into LiteLLM on 25 June 2026. The pull request provides the implementation, review and regression tests.
  • Python
  • pytest
  • Streaming APIs
Read merged pull request (opens a new tab)
Sources & scope

Supplied resume and upstream pull request. Merge status was verified in the source inventory on 12 September 2026.

Further explorations

Trace Zero

A market-replay execution simulator comparing strategies in isolated exchange instances. Listed as lead engineer in the project repository.

View source (opens a new tab)
Browse the wider project archive
TrustSphere Investigator CopilotAn investigator copilot prototype with source-cited evidence and human-approved workflow changes. (opens a new tab)DataLensNatural-language analysis of CSV and Excel data with sandboxed Python execution. (opens a new tab)WayfinderSemantic retrieval and multi-stage verification for four-week upskilling roadmaps. Top 13 at NTU Techfest 2026. (opens a new tab)SpectraAn experimental video-call overlay combining facial-expression and voice-prosody analysis. (opens a new tab)AuroraA modular-agent career development platform built with Next.js, FastAPI, LangChain, and Qdrant. (opens a new tab)Market ReplayTop-of-book capture and replay with monotonic timestamps and configurable replay speed. (opens a new tab)CAPM Portfolio OptimizerCAPM regression, mean-variance portfolio optimisation, and efficient-frontier visualisation. (opens a new tab)AI Engineering from ScratchA learning repository exploring neural networks, transformers, agent loops, and multi-agent coordination. (opens a new tab)PollPulse TNA repository exploring sentiment analysis for Tamil Nadu election data. (opens a new tab)CutTheURLA URL shortener built with Flask and SQLite. (opens a new tab)AdichfyFXA modular browser soundboard with configurable categories and keyboard shortcuts. (opens a new tab)

Research

The Arc

NTU URECA · Aug 2026 — Present

When an agent reviews an experiment, what makes its conclusion trustworthy?

I’m an undergraduate researcher in the NTU URECA programme, advised by Prof. Wang Chen. My work establishes a testbed for studying agent review reliability.

PlannerExperimenterReviewer

What I built

A Python workflow with Docker-isolated experiments, immutable evidence, enforced budgets and Langfuse tracing. The aim is to make an experiment and its review inspectable.

What has been measured

Three measured investigations and a six-episode paired study of evidence handoffs, with independent numerical-claim checks and nine reconciled traces.

This is ongoing research and a working testbed. No published paper or general reliability improvement is claimed.

Sources & scope

Scope, advisor and completed investigations are described in the supplied resume. A public research paper, repository and evaluation report have not been supplied.

Download resume ↓

Experience

Career garden · beside The Hive

From production pipelines to agent reliability.

2025 — now

Follow the milestones. Roles can overlap; dates show when each began.

7 of 8
IndustryAug 2026 — Present

CRPO, NTU

LLM Agent Developer Intern

Making compliance-agent behaviour measurable.

Built a RAGAS evaluation harness covering 200+ compliance-agent test cases across ISO 27001, SOC 2, IM8 and GDPR, on a CRPO-funded project with industry partner Cyber Sierra.

Showing CRPO, NTU: LLM Agent Developer Intern, Aug 2026 — Present.

Read the complete pathway

  1. Feb — Nov 2025 · Industry

    Netvirta

    Software Engineer Intern

    Automated catalogue ingestion for 40+ brands and 100,000+ SKUs. Reduced size-chart processing from 4–5 hours to under 30 minutes per brand, with human review for exceptions.

  2. Aug 2025 — May 2029 · Education

    NTU Singapore

    Computer Science undergraduate

    Bachelor of Computing in Computer Science (Honours) at Nanyang Technological University.

    Degree in progress. May 2029 is the expected completion date.

  3. Dec 2025 — May 2026 · Industry

    eLife Inc

    Backend Engineer Intern

    Built asynchronous ingestion workers processing 10,000+ jobs daily in production. A prompt-tuning harness raised average LLM-judge agreement from 3/10 to 7/10 against ChatGPT web responses. Migrated six scrapers to typed Python.

  4. Jan — Jun 2026 · Leadership

    NTU CCDS TOP’26

    Head of Technology (Backend)

    Led seven backend engineers to deliver an orientation and live-games platform for 900 students, owning frontend integration against a fixed launch deadline.

  5. May — Aug 2026 · Industry

    Cyber Sierra

    AI Engineer Intern

    Built a codebase copilot used by two product managers. Reduced average session cost by 45% from proof of concept to deployment through agentic RAG, caching, parallel subagents and iteration limits. A separate code-cited knowledge-base platform reached internal testing.

  6. 25 Jun 2026 · Open source

    LiteLLM

    Open-source contributor

    Unified prompt-cache usage conversion across streaming and non-streaming Anthropic Messages responses. Added regression tests for cache fields and delayed streaming usage; PR #31076 was merged upstream on 25 June 2026.

    Read the case studyMerged pull request (opens a new tab)

  7. Aug 2026 — Present · Industry

    CRPO, NTU

    LLM Agent Developer Intern

    Built a RAGAS evaluation harness covering 200+ compliance-agent test cases across ISO 27001, SOC 2, IM8 and GDPR, on a CRPO-funded project with industry partner Cyber Sierra.

  8. Aug 2026 — Present · Research

    NTU URECA

    Undergraduate Researcher

    Advised by Prof. Wang Chen. Built a Planner–Experimenter–Reviewer testbed with Docker-isolated experiments, immutable evidence, enforced budgets and Langfuse tracing. Completed three measured investigations and a six-episode paired study of evidence handoffs, with independent numerical-claim checks and nine reconciled traces.

    Ongoing research and a working testbed. No published paper or general reliability improvement is claimed; a public research repository or evaluation report has not been supplied.

    Explore the research
Download full resume ↓

About

Yunnan Garden

AI engineering · Software engineering

I’m a Computer Science undergraduate at NTU. I build agentic tools, evaluate their reliability, and engineer the systems behind them.

I work across applied AI and backend engineering: turning an agent idea into a usable workflow, then checking whether its behaviour is supported by evidence.

Education

Bachelor of Computing in Computer Science (Honours)
Nanyang Technological University
Aug 2025 — May 2029 · In progress

Tools I work with

Languages
Python, TypeScript, JavaScript, SQL
Agents & evaluation
LangChain, LangGraph, DeepAgents, MCP, agentic RAG, RAGAS, Langfuse
Backend & infrastructure
FastAPI, PostgreSQL, pgvector, Redis, Temporal, AWS, Docker, Kubernetes, Terraform

Why a campus?

NTU is the setting for my studies and research. This original, stylised campus turns the portfolio into a small place to explore, while keeping every section one click away.

An independent personal portfolio. This project is not affiliated with or endorsed by NTU. Geography and buildings are intentionally simplified.

Contact

Campus welcome

Let’s talk about useful AI and well-built software.

For roles, project questions or a conversation about my work, email is the most direct way to reach me.

LinkedInProfessional profile (opens a new tab)GitHubCode, projects and contributions (opens a new tab)Resume ↓PDF · Professional experience