nelson@imperial:~

$cat profile.txt

Nelson Gong

[2nd Year MEng Mathematics & Computer Science]

[Imperial College London]

[John and Rita Henke Scholar · £12.5k p.a.]

[Amazon SDE Intern · AWS Security · Summer 2026]

I love building things. Currently building knownissue.dev: shared memory for AI coding agents. Specialising in MCP and AI agent tooling.

Previously at BlackRock as a Software Engineering Spring Intern. IC Algothon competitor and IBM Z Datathon 2nd place winner.

$ls -la ~/featured# agent infrastructure: memory, governance, knowledge

~/knownissue.dev

live

Shared memory for AI coding agents

MCP server that coding agents query when they hit errors. Implements a four-verb loop (search, report, patch, verify) where verification is binary and automatic: does the failing repro now pass? Listed on Glama’s MCP registry; YC referenced it as a promising application.

Why it matters

Most bug-fix knowledge dies in the conversation where it was solved. Verification is the moat: it keeps data quality high at scale without human-in-the-loop curation.

TypeScriptHonoNext.jsPrismaPostgrespgvectorOAuth 2.1MCP

→ live ⌥ repo

~/WACC Compiler

completed

Full Scala compiler targeting AArch64

End-to-end compiler for the WACC language: lexing, parsing, semantic analysis, code generation. Compiles to AArch64 assembly, cross-compiled with aarch64-linux-gnu-gcc and run under qemu-aarch64. Built for Imperial COMP50007 as a group project.

Why it matters

Designed with a Stack IR layer between AST and backend so an optimisation pass can slot in cleanly, rather than locking the codegen into a pure stack machine.

ScalaparsleyStack IRAArch64

⌥ repo: private (Imperial GitLab)

~/Willow

beta

Multi-agent knowledge-graph memory for AI

Consumer chat agent built on a structured, evolving knowledge graph of user information. ~20 active beta testers. The real product is the memory layer underneath; the chat is just the surface that exercises it.

Why it matters

You can’t bolt a "remember this" tool onto a chat LLM and expect good memory. Willow runs separate background agents for indexing, graph evolution, maintenance, and retrieval, because the chat-facing model is too lazy about upkeep.

TypeScriptMulti-agentKnowledge Graph

⌥ repo

~/CramKit

completed

Local-first AI study tool with an MCP-first interface

Upload exam materials → converted to markdown → chunked → indexed → exposed to Claude Desktop via MCP. Studying happens inside Claude; the web UI exists only for data management. SQLite plus local filesystem, no cloud dependencies. Dogfooded on real exam prep.

Why it matters

MCP-first design treats Claude as the primary user. The web app exists only to manage the data the agent will read. A useful pattern for tools where the model is the consumer.

pnpmViteReact 19HonoBunPrismaSQLiteMCP SDK

⌥ repo

~/Sentinel

archived

MCP control plane for organisations deploying AI agents

Authorization and governance layer for MCP: server registration, RBAC policies, encrypted credential proxying, audit logging, OAuth 2.1 proxy for upstream MCP servers. ~500k LOC, deployed at sentinel.london. Sunset after well-funded competitors (Runlayer, golf.dev) entered the same space.

Why it matters

Direct lineage to knownissue: both MCP infra, and Sentinel’s auth model informed knownissue’s. Killing it cleanly when the market moved was the right call.

HonotRPCPostgresPrismaOAuth 2.1React

⌥ repo: private

$ls -la ~/earlier-work

~/Crucible Trader

completed

Modular quantitative trading platform with backtesting engine, statistical validation (permutation tests, walk-forward analysis), and C++ pricing library. Built for rigorous hypothesis testing of trading strategies. Local-only deployment.

TypeScriptNext.jsC++20PythongRPCDuckDBParquet

~/PintOS

completed

Extended Imperial's PintOS educational operating system. Implemented priority scheduling with priority donation, MLFQS scheduler, user program execution with stack setup, system calls, and parent-child process synchronisation.

Cx86 AssemblyOS DesignConcurrency

~/ARM Assembler & Emulator

completed

Two-pass assembler supporting 50+ AArch64 instructions with label resolution. Full ARMv8-A emulator simulating pipeline stages, 31 GP registers, 2MB memory, and GPIO control for Raspberry Pi 3B LED circuits.

CARM AssemblyGitLaTeX

~/Texas Hold'em Poker Trainer

live

Interactive poker game with AI opponents featuring Monte-Carlo equity simulation, ICM tournament analysis, GTO range lookups, and real-time EV calculations. Includes configurable bot personalities, advanced hand analysis, and a standalone practice tool.

ReactTypeScriptWebpackGame AIMonte Carlo

~/ML Fraud Detection

completed

Machine learning model for financial fraud detection achieving 97.68% accuracy. 2nd place in London, winning £250, IBM LinuxOne server access, and 3-month mentorship. Processed large-scale transaction data on IBM Z platform.

PythonScikit-learnTensorFlowIBM ZData Analysis

~/Derivative Calculator

live

Symbolic differentiation engine with step-by-step explanations and interactive graphing. Features dynamic tangent line visualisation and real-time expression parsing. Achieved full marks (70/70) in A-level NEA.

Next.jsTypeScriptReactTailwindCSS