Will Zhang will03216

class WillZhang:
    role      = ["Forward Deployed Engineer", "Solutions Architect", "Software Engineer"]
    education = "MEng Electronic & Information Engineering @ Imperial College London (2027)"
    shipped   = {"BytePlus": "Solution Architect Intern — AI agents, MCP, GenAI video",
                 "ByteDance": "iOS Software Engineer Intern — Swift / Objective-C"}
    stack     = ["Python", "TypeScript", "C++", "Swift", "SystemVerilog"]
    languages = ["English", "Mandarin"]

I work best between the customer and the code. At BytePlus I sat with sales teams and customers, found where their workflow broke, and shipped the fix — a lead-qualification engine, internal agents, an MCP server and a full video studio. Before that I shipped production iOS features at ByteDance. My roots are in systems: I've built a C compiler and a pipelined RISC-V CPU, so I'm comfortable going all the way down the stack when a problem needs it.

💼 Experience

Solution Architect Intern · BytePlus (ByteDance), London _{Mar – Sep 2026}

Built a lead-generation engine that replaced manual lead qualification: intent signals → LLM scoring → a multi-agent system where specialist agents compete for the send budget. Knowledge-base-linked outreach lifted the reply rate by 16%.
Worked directly with customers on generated-video defects (identity drift, grid artefacts), traced them to diffusion-model behaviour and fixed them through prompt, framing and resolution changes.
Shipped a Seedance MCP server (SDK + server + CLI + agent skills) so non-engineers could generate video from their AI assistant, deployed once in Docker with server-side credentials.
Shipped three internal agents for Sales / AE teams: a RAG knowledge-base agent, a pre-call prospecting agent and a Slack daily digest.
Built Lumen Hub, an infinite-canvas AI video studio in Next.js / React.

iOS Software Engineer Intern · ByteDance, Beijing _{Jul – Oct 2025}

Built comment-input pre-creation and reuse infrastructure (object pooling, prewarm scheduling): −20–27 ms first render, +0.2–0.48% comment clicks per user.
Added a Swift / Objective-C pre-commit quality pipeline (SwiftLint, OCLint, Git hooks) and migrated 4 UI components from Objective-C to Swift.

🧠 Featured Projects — AI Agents & Customer-Facing Systems

Lead Qualification Engine

A "Claude brain" for outbound sales — turns noisy lead exports into a ranked, auditable lead list.

Two-stage LLM pipeline: evidence extraction → scoring, separating observed facts from inference
Python control plane: score clamping, threshold routing, hard disqualifiers
Per-lead JSON audit artifacts · FastAPI UI · Postgres · Docker
Provider-agnostic (Claude / Seed 2.0 via .env)

Lumen Hub — AI Video Studio

A self-hostable infinite-canvas studio for Seedream / Seedance generation.

An LLM splits one concept into shots under the model's length limit; ffmpeg stitches them server-side into one long video, and interrupted runs resume where they stopped
Video Style Extract: reference clip → style brief → enhanced prompt
Strict BYOK: per-user encrypted API keys, Clerk auth, Postgres
Dockerised standalone Next.js build with health checks

Pre-Meeting Customer Profile Agent

Writes account briefings for AEs straight from Lark calendar events, domains or emails.

Scans calendar windows, skips internal meetings, auto-detects the customer's vertical
Apollo.io + web enrichment with domain-identity guards against false matches
Evidence-labelled answers: Known / Likely / To Validate / Not Found
Industry playbooks · keeps working when optional permissions are missing · packaged as a portable skill

Slack Daily Digest

An agent skill (Claude Code / Codex / Mira) that turns Slack into a twice-daily AI intelligence digest in Lark Docs.

Executive summary, TODOs, pipeline, risk signals, per-channel transcripts
Self-healing watchdog: validates output, retries with backoff, escalates
Local scheduler (launchd / systemd) that catches up on runs missed while the machine was off

🔌 MCP Servers & SDKs

Project	What it is	Highlights
Seedance MCP	Video generation for Claude Desktop, Cursor, Windsurf & the terminal	SDK + MCP server + CLI + Claude Code skill · submit-then-poll jobs with cancel & history · automatic asset upload for portrait references · edit/extend · checks requests are valid before submitting
Seedream MCP	Image generation across Seedream 4.5 / 5.0 Lite / 5.0 Pro	Zero-dependency SDK · layer decomposition into RGBA cutouts · TOS upload with HMAC signing · CI on Py 3.10–3.12

⚙️ Systems & Hardware

Project	What it is	Stack
C → RISC-V Compiler	Compiles C with pointers, scoped variables, control flow and function calls to RISC-V assembly. Saves live registers to the stack across calls so nested expressions don't overwrite each other	C++ · Flex/Bison · RISC-V
Pipelined RISC-V CPU	RV32I, 5-stage pipeline, hazard unit, 2-way set-associative cache — team of 4	SystemVerilog · Verilator · FPGA
FPGA Bop It	Multiplayer game on DE10-Lite + NIOS II with an AWS game server · demo	Verilog · C · Python · AWS
Self-Balancing Robot	Two-wheeled robot with PID control on ESP32 + MPU6050, Raspberry Pi link	C++ · Python · Embedded

🛠 Tech Stack

Agentic AI & LLM: multi-agent systems · MCP · agent skills · RAG / embedding retrieval · prompt engineering · CrewAI · AgentOps · OpenAI API (GPT, Whisper) · Claude API · video / image / 3D generation APIs
Engineering: Python · TypeScript · C++ · C · SQL · Swift · Objective-C · Bash · SystemVerilog · WebSockets · OAuth · ffmpeg

_{⭐ The repos above are the best place to start. Happy to walk you through any of them.}

Navigation Menu