// kpriyadharshan.dev
Kannan
Priyadharshan
AI Engineer · Agentic Infrastructure
I build agent systems that run against production code and real users — retrieval that has to be correct, guardrails that have to hold, evaluation that tells you when they don't. CS @ NTU 2029. AI Engineer @ Cyber Sierra.
02 · Selected Work
Projects that
mean something
01 · AGENTIC AI
shopee-live-producerLive-commerce agent that behaves like a producer, not a chatbot — it classifies each buyer comment and decides whether to auto-answer, escalate to the host, flag policy risk, or stay silent. A grounding gate blocks any reply not backed by product-fact IDs above a confidence threshold, and host-confirmed answers extend session memory without overwriting the seeded catalogue. 3rd of 40 teams at the Sea × OpenAI Codex Regional Hackathon 2026.
Next.js · LangChain DeepAgents · OpenAI · Supabase Realtime · TypeScript
View on GitHub →02 · AGENTIC AI
wayfinderAI career navigation engine built in a sub-24-hour sprint. A multi-stage agent pipeline combines pgvector semantic retrieval with LLM re-ranking and a verification layer, then generates personalised 4-week upskilling roadmaps. Top 13 / 100+ teams at NTU Techfest 2026.
Next.js · pgvector · OpenAI · PostgreSQL
View on GitHub →03 · OPEN SOURCE
litellm · PR #31076Merged contribution to LiteLLM, the 52k-star gateway that routes calls across 100+ LLM providers. Centralised the OpenAI → Anthropic usage translation so prompt-cache token counts survive every response path in the Anthropic messages adapter — cache hits were being dropped from usage reporting, which quietly breaks cost accounting for anyone billing on cached tokens.
Python · Anthropic API · OpenAI API
View merged PR →04 · NLP
pollpulse-tnReal-time NLP sentiment pipeline for Tamil Nadu 2026 election forecasting. Aggregates and classifies social signals across sources to surface swing-district indicators.
Python · NLP · Supabase
View on GitHub →05 · SYSTEMS
trace-zeroReal-time execution simulator built on nanosecond-accurate orderbook capture. Every strategy runs against its own isolated exchange instance, so one lane’s price impact can never contaminate another; monotonic clocks are the authoritative timing source, and state streams to the front end over WebSocket. The domain is optimal execution (Almgren-Chriss, 2000) — the engineering is deterministic replay and state isolation.
Python · FastAPI · Next.js · WebSocket · NumPy
View on GitHub →06 · SYSTEMS
market-replayNanosecond-accurate top-of-book capture and replay. Monotonic clock timestamps — immune to DST and NTP adjustments — as the authoritative timing source. Pluggable handler system, JSONL storage with 10-minute file rotation.
Python · WebSocket · Binance
View on GitHub →07 · QUANT
capm-portfolio-optimizerBeta estimation via linear regression against a market index. Constructs the optimal portfolio using Markowitz Mean-Variance Optimisation. Monte Carlo simulation across 10,000 weight combinations to map the efficient frontier and identify the max Sharpe ratio portfolio.
Python · NumPy · SciPy
View on GitHub →03 · Experience
Where I've
shipped
AI Engineer Intern
- Built CKA, an internal agent platform that answers natural-language questions over production codebases and runs packaged agent skills to surface hidden edge cases — collapsing product-planning turnaround from multi-hour engineering round-trips to self-serve sessions
- Cut average trace cost between POC and production at sub-minute latency by designing retrieval around agentic RAG — navigating repositories on demand rather than pre-indexing — with parallel subagents, prompt caching and iteration caps that degrade gracefully
- Built the customer-facing knowledge-base platform: assembles product guides from engineering wikis with code-cited evidence, gates publication behind automated misleading-content checks, and answers end users through a guardrailed chat agent that disclaims or refuses rather than speculates
- Assessed adversarial prompts for harm feasibility and exploitability against an internal rubric in an AI safety evaluation
Backend Engineer Intern
- Raised API-to-web-UI response agreement from 3/10 to 7/10 by building a prompt-tuning harness that iteratively refined a seed system prompt against ground truth, using an LLM judge for semantic similarity and style
- Scaled ingestion to 10,000+ daily jobs across YouTube, X and Reddit by designing a unified async worker base on AWS SQS with exponential-backoff retries and centralised error handling
- Ported 6 source scrapers and their pipelines from TypeScript to Python under mypy, and measured client product visibility in generative search using Temporal-orchestrated probe workers
Software Engineer Intern
- Automated merchant catalogue onboarding for 40+ brands and 100k+ SKUs by building Python pipelines that ingest product data from regional storefronts and classify attributes to power tagging in a social-commerce product
- Cut per-brand size-chart processing from 4–5 hours of engineer time to under 30 minutes by shipping an LLM extraction tool that converts merchant product links into Pydantic-validated JSON at 70–80% automation, with human-in-the-loop fallback
Head of Technology (Backend)
- Delivered the platform powering freshman orientation and its live games for 900 students against a fixed launch date, coordinating 7 backend engineers of mixed experience to sprint milestones
- Owned the frontend integration that gated launch, plus code review and deployment scheduling across the team
04 · At Work
What I build
in private repos
Most of my production work lives behind a company login, so here is what it actually does.
Agent infrastructure at Cyber Sierra. I build the internal agent platform that lets the product team query production codebases in natural language and run packaged agent skills to surface edge cases hidden in the code. Retrieval is agentic rather than a pre-built index — the agent navigates the repository and pulls code on demand, which avoids an index that goes stale on every commit. The work I care about most there is the unglamorous part: driving down cost and latency per trace between prototype and production without losing answer quality.
Customer-facing knowledge base. I also built the platform that assembles product guides from engineering documentation with code-cited evidence, gates publication behind automated misleading-content checks, and answers end users through a chat agent constrained to disclaim or refuse rather than speculate. Refusing well turned out to be harder than answering well.
Happy to walk through any of it in detail.
05 · Stack
Tools I
know well
06 · Writing
Thinking
out loud
Two of Our Three Demos Failed Live. We Still Placed Third.
We built Shopee Live Producer — an agent that classifies buyer messages, grounds every answer against a verified catalogue, and decides what actually reaches the customer. Two of three demos broke on stage at the Sea × OpenAI Codex Regional Hackathon. What survived was the part the judges could follow: a coherent journey and legible core logic.
Read on Medium →07 · About
Who I
am
I build agent systems that run against production code and real users — retrieval that has to be correct, guardrails that have to hold, and evaluation that tells you when they don't. Right now that means agent infrastructure at Cyber Sierra, an AI-powered GRC platform in Singapore.
Before that: prompt-tuning harnesses and high-throughput ingestion at eLife — an LLM judge scoring response agreement, and an async worker base moving 10,000+ daily jobs. At Netvirta: an LLM extraction tool that took per-brand size-chart processing from 4–5 hours of engineer time to under 30 minutes, alongside catalogue pipelines onboarding 100k+ SKUs.
At core, I'm a problem solver. The domain matters less than the depth — I started in low-latency market infrastructure, and the same instincts (determinism, isolation, measuring the thing rather than guessing at it) are what I now bring to agents.
Outside of this: varsity cricket for NTU, Hiphop Tamizha, and a long-term plan to own a coffee estate in Kodaikanal.
Leadership
Reading
Get in Touch
Open to internship opportunities for Summer 2027 in Hong Kong and Singapore.
Particularly interested in agentic AI infrastructure, backend systems, and applied AI research roles. If you're building at the intersection of AI and real-world systems, I'd like to hear about it.