Skip to content
View Dharshan2004's full-sized avatar

Highlights

  • Pro

Block or report Dharshan2004

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Dharshan2004/README.md

Kannan Priyadharshan (Dharshan)

AI Engineer · Agent Systems & Evaluation
CS @ Nanyang Technological University · Expected 2029
LLM Agent Dev Intern @ Cyber SG R&D Programme Office


What I'm building

  • Plexus — I'm building a collaborative workspace where teammates can supervise coding agents together, share sessions, approve actions, and review changes. Currently developing the desktop experience and onboarding.
  • Agent evaluation @ Cyber SG R&D Programme Office — working on evaluating agent behaviour and evidence grounding, including how agents handle incomplete or conflicting information.
  • Agent infrastructure @ Cyber Sierra — built tools for querying production codebases and a customer-facing knowledge agent that uses code-cited evidence and checks generated guides before publication. Refusing well turned out to be harder than answering well.

What I care about

I build agent systems and the tools to understand where they fail. I'm interested in how people supervise them, how they use evidence, and whether an evaluation actually tells us what to improve.

My earlier work includes high-throughput data pipelines at eLife (AWS SQS, gRPC, Kubernetes), ETL infrastructure at Netvirta (AWS EKS, 100k+ SKU onboarding), and a document ingestion pipeline for Skills@CCDS with schema-constrained extraction and human verification.

I started in low-latency market infrastructure. Deterministic replay, isolated state, and measuring behaviour are still useful instincts when working with agents.


Research

  • NTU URECA — building and evaluating scientific-agent workflows, with a focus on reliability, evidence quality, and resource use.
  • Algoverse — exploring shared blind spots in multi-agent safety evaluation and how evidence selection affects reviewers' judgments. Research in progress.

Projects

Project What it does Stack
Plexus (private) Collaborative agent workspace with shared sessions, teammate approvals, handoffs, and change review. Under development. JavaScript · Electron · Node.js · WebSockets · SQLite
shopee-live-producer Live-commerce agent that routes buyer comments to grounded answers, host escalation, or policy review. 🥉 Sea × OpenAI Codex Regional Hackathon 2026. Next.js · LangChain DeepAgents · OpenAI · Supabase Realtime
wayfinder Maps résumé skills to target roles and builds a four-week learning roadmap with linked resources. Top 13 / 100+ teams at NTU Techfest 2026. Next.js · pgvector · OpenAI
trace-zero Execution simulator with deterministic orderbook replay and isolated exchange instances for comparing strategies. Python · FastAPI · Next.js · NumPy
Earlier work
Project What it does Stack
pollpulse-tn NLP sentiment pipeline aggregating social signals for Tamil Nadu 2026 election analysis. Python · NLP · Supabase
market-replay Top-of-book capture and replay with monotonic timestamps, pluggable handlers, and rotating JSONL storage. Python · WebSocket · Binance
capm-portfolio-optimizer Beta estimation, Markowitz portfolio optimisation, and Monte Carlo simulation across 10,000 weight combinations. Python · NumPy · SciPy

Open source

  • LiteLLM #31076 — merged. Fixed missing cache-read and cache-creation token counts in the Anthropic messages adapter by sharing usage translation across streaming and non-streaming responses.

Stack

Agents      LLM Orchestration · Multi-Agent Systems · Agentic RAG · Evaluation
Languages   Python · TypeScript · JavaScript · Go · C++
Backend     FastAPI · Node.js · gRPC · WebSockets · Redis · PostgreSQL
Cloud       AWS (EKS, SQS) · Kubernetes · Docker · Terraform
ML          OpenAI Embeddings · LangChain · Qdrant · pgvector

Writing


On my bookshelf

  • Nassim Taleb — Antifragile
  • Bhagavad Gita, Chapter 14 — on the three modes of nature

Beyond the terminal

My campus work includes backend leadership for NTU CCDS TOP'26, with a team of 7 engineers building a platform for 900 incoming students, and directing the Quantitative Finance Academy at NTU, with curriculum on market microstructure and ML in finance for 80+ members.

Varsity cricket. Long-term goal: a coffee estate in Kodaikanal.


📍 Singapore → Hong Kong (HKU exchange, Jan 2027) · Open to Summer 2027 roles

🌐 kpriyadharshan.dev  ·  💼 linkedin.com/in/kpriyadharshan  ·  📧 kpd2204@gmail.com

Pinned Loading

  1. shopee-live-producer shopee-live-producer Public

    🥉 3rd place — SEA x OpenAI Codex Regional Hackathon 2026. AI-powered livestream producer demo. Sellers build a product lineup | A LangChain DeepAgent auto-replies to buyer comments with grounded pr…

    TypeScript 5 3

  2. wayfinder wayfinder Public

    AI Career Navigation Engine: Transforming fragmented career data into actionable 4-week upskilling roadmaps using pgvector semantic search and multi-stage LLM verification. Top 13 Finish (Honourabl…

    TypeScript 1

  3. trace-zero trace-zero Public

    AC optimal execution simulator with market microstructure replay, permanent/temporary price impact, and multi-strategy tear sheet analysis. Full-stack backend (FastAPI) + Bloomberg Terminal-style f…

    Python 1

  4. market-replay market-replay Public

    Lightweight Python Market-data CLI to normalize & store top-of-book, then replay with timing for execution backtests

    Python 1

  5. capm-portfolio-optimizer capm-portfolio-optimizer Public

    A Python-based quantitative tool that estimates asset Beta via linear regression and constructs an optimal portfolio using Mean-Variance Optimization (Markowitz). Features Monte Carlo simulations t…

    Python 1

  6. Aurora Aurora Public

    Aurora is an AI-powered career growth platform that uses modular agents for onboarding, skill navigation, and progress tracking. Built with Next.js, FastAPI, LangChain, Qdrant, and GPT-4 to deliver…

    Python 2