Wenhui (Kit) Kang

Independent quant and MEV developer

Low-latency trading systems in Rust, Go and Python, Vancouver

• Built a Go gRPC bundle relay, later ported to Rust (tokio, axum), fanning Jito bundles to 8 block-engine regions over 4,000+ egress IPs with token-bucket limits and Retry-After cooldowns.
• Fixed a relay throughput collapse with a shared atomic limiter and a min-heap ready queue: proxy wait fell from 13–28 ms to under 0.5 ms.
• Built a Rust Solana atomic-arbitrage bot peaking at about $80M daily volume, with Jito bundles, Kamino flash loans, compute budgets for 22 DEXs and an on-chain profit guard.
• Wrote about 80% of a live Rust Polymarket trading system (720 of 894 commits): Binance SBE feed, signals, execution, risk, reconciliation and gasless claims.
• Hand-wrote its Binance SBE decoder, about 1.65 s ahead of the Chainlink feed that settles each market; priced markets as binary options (N(d2)) behind eight entry gates; validated on a live-parity backtester over 17,549 markets with sweeps of up to 100,800 configurations and out-of-sample checks.
• Built Robinhood Chain (Arbitrum L2) arbitrage with Bellman-Ford cycle search, bit-exact Uniswap V3/V4 tick math and Solidity flash executors; sign-and-send went from 672 ms to 12 ms.

Solana relay case study →Polymarket case study →

Co-founder, product manager and advisor

RootData, crypto data platform

• Co-founded the platform and designed and collected its database from zero; as PM grew it to 428K users (about 14× organic in 12 months), 700K–1M daily page views, 22K+ projects and 200+ API partners including Binance Alpha and OKX Boost.
• Built data charting, SEO, anti-crawling, UI and CRM; open-sourced TypeScript and Python MCP servers exposing about 20 RootData tools to AI agents.
• Built a 61-model XGBoost analytics app over 6,000+ projects (launch AUC 0.81) from 116 automated experiments.

RootData case study →

Website project manager and full-stack developer

Telecom reseller (client), shown here as WestTelecom

• Shipped a bilingual plan catalogue on Next.js 13 App Router, TypeScript and Sanity v3 across mobile, internet, TV and security (229 commits).
• Built filtering, sorting, comparison and modal UIs; wired HubSpot lead capture, NextAuth and signed webhooks for on-demand revalidation; later audited it and moved leaked secrets server-side.

WestTelecom case study →

Investment advisor

J17 Crypto, family office (remote)

• Advised portfolio management and due diligence for about $90M in digital assets; built Dune SQL dashboards for position and risk monitoring.

Investment partner

CatcherVC

• Reviewed 200+ seed deals a year and managed $4M+ in digital assets; researched ZK proofs, account abstraction and FHE/MPC. Portfolio includes MegaETH, GRVT and Scroll.

Overseas operations and business development

ChainCatcher

• Co-interviewed founders of ether.fi, LayerZero Labs and Wormhole Foundation; co-hosted a Hong Kong Web3 Festival 2023 side event; closed BD deals with LayerZero, Mina and Consensys.

LLM-Digest, self-hosted market intelligence

Next.js 16, React 19, TypeScript, AI SDK v6, SQLite FTS5

• Nine-stage narrative LLM pipeline with Zod-validated JSON hand-offs, jsonrepair and retry, byte-level stall watchdogs and per-task model fallback; 4,100+ reports from 127 sources.
• Tool-calling RAG agent (6 tools, phased 3-step loop) fusing SQLite FTS5/BM25 with 12 web search sources by reciprocal rank fusion, with query rewriting and citation checks.
• Semantic news dedup with NVIDIA embeddings, SimHash and cosine-centroid clustering over in-memory Float32 pools, chosen over a vector database after benchmarking Qdrant and Postgres.
• One client over 7 LLM providers: encrypted key vault, per-provider pacing, daily budgets, tool-call repair and per-run cost tracking.
• Self-hosted with systemd and Cloudflare Tunnel; staged-build deploys cut downtime from about 2.5 minutes to about 3 seconds.

LLM-Digest case study →

KV-cache coordinated LLM serving, graduate capstone

Python, vLLM, Triton, Llama-3.1-8B on an RTX 3090; team of 4

• Owned the KV-cache layer under a learning-to-rank vLLM scheduler: 4-bit KV gave 2.83× token capacity and cut mean time to first token 6.8× at 64 requests a second.
• Wrote block-sparse attention Triton decode kernels (1.33–1.75× standalone) and patched vLLM's decode path: −22% time per output token, −47% P99.

KV-cache case study →

Stablecoin card-payment indexer

TypeScript, ClickHouse, Postgres, EVM and Solana adapters

• Rebuilds stablecoin card-payment volume from self-indexed chain data, issuer by issuer: 22 of 24 card programs traced, 413 registered addresses each with its evidence, within ±0.6% of the public benchmark on EVM mainnets.

Payment indexer case study →

Local LLM inference lab

Ollama, llama.cpp (CUDA), GGUF K-quants, ComfyUI, Hugging Face

• Run and tune quantized 27B LLMs and vision models on one 24 GB RTX 3090; trained and published an OPT-125M ranker and dataset on Hugging Face.

Notes on a 27B model on one RTX 3090 →

MSc, Applied Computer Science

Fairleigh Dickinson University, Vancouver

• CGPA 3.56/4. Capstone, Distributed Database Systems and Advanced Operating Systems: A.

BBA, Finance

Beedie School of Business, Simon Fraser University

• CGPA 3.4/4, Statistics 3.8/4. Invited to the UC Berkeley Haas summer program.