AI application / Agent development · Monash AI Master's (QS 37) · Class of 2027
Huang Yihang
Turning LLMs into Agents that actually get things done.
A fix PR merged into an overseas open-source AI project · enterprise Q&A at 85%+ accuracy · an open-source desktop AI assistant that operates your computer locally — from memory architecture and safe sandboxing to tool calling, I build Agents that ship and hold up to scrutiny.
AI twin · tap to start chatting
The headlines first.
Say it in one sentence; your computer does the rest.
NoWorries · solo build · download from the official site · files aren't uploaded to the cloud
In productionA staff member types one line in a group chat, and the order and stock check are done.
Chenxi Flowers · solo build · RMB 5.4M+ in cumulative company sales · 5 Feishu group AI Agents · the AI never gets database write access
Commercial productUpload a résumé, and 8 Agents generate a differentiated personal site.
PowerMyWeb · live with Free / Pro subscriptions · self-built generation pipeline · 130+ unit tests, 352 assertions
AI × LawRead a 348-page property contract into one verifiable risk report.
ContractLens · benchmarked against a real lawyer report · 91 cited findings · ~$1 per contract
Open sourceMy fix was merged into an active open-source AI project.
OpenClaw · PR merged · also rebuilt its memory and context modules
Enterprise internshipBuilt an AI review pipeline in my internship — 3 hours of manual checking down to 10 minutes.
Sugon · Agent dev internship · enterprise knowledge Q&A at 85%+ accuracy
Open frameworkWhen one AI can't finish a large task, it grows a task tree.
Sprout · MIT open framework · same capped-token task: single AI 25, Sprout 100
Flagship open-source project · solo build · MIT, free and open
NoWorries NoWorries
Say it once. Leave the rest to it.
Give it one instruction, and it works locally across your files, Excel, and Word.
A deep rebuild on open-source Openwork — Electron + TypeScript + Python, one person carrying the full stack.
Files never leave your machine. Every step can be rolled back.
Memory system
Three tiers of memory. It gets you, the more you use it.
Local SQLite · vector embeddings + semantic search · incremental summarization · emotional tagging
It can only touch what you allow.
Allowlisted workspace isolation, dangerous-command blocking, sensitive-path protection, and a full operation log — on macOS it's a real OS-level sandbox (sandbox-exec, deny-all by default). Even first-timers can hand it real work.
Use your own key, no subscription fee — Zhipu GLM · Kimi · SiliconFlow, or any OpenAI-compatible endpoint (DeepSeek / Ollama both work)
Ask the AI twin how the three-tier memory was designed ›Solo build · running in real production
The whole business runs in a group chat.
The operations system a flower wholesaler uses every day: a staff member types one line in a Feishu group, and the order and the outbound stock check are done.
Chenxi Flowers runs on it every day: 3 growing sites, 65 varieties, 56 customers, 10 staff, and RMB 5.4M+ in cumulative company sales; 1,027 orders and RMB 340K+ settled in the first month live. It started as a Feishu Bitable MVP to prove the business loop out, then — once the business outgrew it — was rebuilt into a Flask + PostgreSQL production system that has run steadily ever since.
Zhang's flower shop needs 200 bunches of grade-A red roses — afternoon truck, ship first
AI has drafted the dispatch note
Feishu Bitable MVP → Flask + PostgreSQL · 30+ permission-controlled endpoints · daily off-site backups
Knowing when to change the skeleton.
A Feishu Bitable MVP stood up in a few days proved the business loop out and validated what was really needed; once the business grew to 3 sites, 65 varieties and 56 customers, Bitable stopped holding up on concurrent writes, reconciliation consistency and complex queries — only then was it rebuilt on Flask + PostgreSQL. The hard part was never writing the system; it's judging when to change the architecture.
The AI never gets write access.
Read-only by default, writes through an allowlist, execution only after human confirmation — the AI only drafts the note, the system is what commits it, and the AI never touches the database's write permissions. Getting an owner comfortable handing over order entry and account checks is the hardest step for an Agent entering real production.
Not a demo — it's used every day.
10 staff run the business on it daily, with RMB 5.4M+ in cumulative company sales and 1,027 orders and RMB 340K+ settled in the first month live. Delivered to production standards: 30+ permission-controlled endpoints, full operation auditing, daily off-site backups, and automatic rollback on failure.
In production · RMB 5.4M+ in cumulative company sales · 1,027 orders / RMB 340K+ settled in the first month · 5 Feishu group AI Agents · zero AI database write access
See the full system and how it evolved ›Ask the AI twin how you get an owner to hand over control ›Live · multi-agent generation engine
I hand-wrote this website. The next one, AI writes for you.
A website builder for job seekers: the AI reads your résumé, talks it through with you like a bespoke tailor, then writes you a one-of-a-kind site from scratch.
PowerMyWeb is officially live, with Free / Pro subscriptions open: upload your résumé, and 8 specialized Agents read your background, mine an inspiration library distilled from 1,806 real portfolios, and draft 5 design directions deliberately pushed apart; after a chat to confirm, they hand-write your site line by line and publish it to your own subdomain in one click.
No LangChain.
Structured output, truncation-resume, tool loops, an NDJSON event stream, a failure-escalation chain — all built from scratch, because I want full control over every step.
Not looking alike is a sampling problem.
The five directions an LLM drafts collapse into sameness — so Verbalized Sampling for explicit sampling + Max-Min diversification, oversampling 8 to pick 5, pulls diversity back from the probability tail. This sampling pipeline is already running live with the product.
AI wrote it, AI reviews it first.
Each version is screenshotted with Playwright and scored by a vision model; changes go through three-tier triage — sub-second patch, block rebuild, full-page fallback. Once a site is live, you can still revise it by chatting and roll back at any point.
Live with Free / Pro subscriptions · 130+ unit tests and 352 assertions · the multi-agent generation pipeline, anti-sameness sampling and visual review loop are all proven out
See the full architecture and progress ›Ask the AI twin how it stops AI designs from looking alike ›ContractLens · benchmarked against a practising lawyer's real review report
Read a 348-page contract down to one clear page.
Property contracts run hundreds of pages and nobody reads them all: this puts 7 AI analysts on the job, each reading closely, and turns every risk into a one-page report you can trace back to the source.
AI review of Victorian (Australia) property contracts — a 10-stage pipeline, 7 AI analysts in parallel; a 132-page contract runs end to end in 4.6 minutes for about $1.
text extraction
扫描页 → Tesseract OCR
Every conclusion holds up to scrutiny.
Mandatory source citations + tiered rapidfuzz verification (85 / 75 / 60 thresholds) + targeted retries only on failed citations; output then passes two compliance gates (regex + AI semantic review), intercepting “AI-lawyer” overreach.
Benchmarked against real lawyer reports.
4 real contracts (104–348 pages, including scanned OCR and a 5-address mixed-title case) ran end to end. Compared line by line against a practising lawyer's report: about 30 findings overlapped. Plus 2 ACN discrepancies the lawyer's report hadn't itemized — confirmed by line-by-line check. The 86 items the system flagged as “needs human review” were walked through one by one too: 85% held up as useful.
Open-source contribution · OpenClaw
Gave OpenClaw's memory a fallback — and a way to forget.
OpenClaw is an overseas open-source AI assistant with 380k+ GitHub stars: one fix PR was merged upstream; a forked version adds a forgetting episodic layer and multi-tier retrieval fallback to its memory.
Retrieval degrades tier by tier with a fallback when it breaks; then an episodic memory layer on top — consolidated by spaced repetition, forgotten when stale.
spaced-repetition
decay ~10%/day「用户偏好用 pnpm」
↑ 升级进 MEMORY.md
An episodic layer on top of hybrid retrieval
Upstream retrieval is already hybrid (vector + BM25); I added an episodic memory layer — each task's result, tools used, and files touched are logged to a monthly JSON, giving retrieval one more fallback path.
Memory that forgets: spaced-repetition consolidation
Importance decays ~10%/day — memory forgets by default; only items recalled ≥3 times within 7 days get promoted into a long-term MEMORY.md. The more you use it, the better it sticks.
Open framework · designed solo
Sprout — a task tree that grows itself.
An open framework that lets AI build its own team: when a job is too large for one call to finish, it splits the work into a task tree and stitches the results back together.
Every Agent decides its own splits — a tree topology that grows recursively, its depth decided by the task.
Root
analyze()
Child A
execute()
Child B
analyze()
Child C
execute()
B-1
execute()
B-2
execute()
Phase 1
analyze()
只判断：要不要拆
Phase 2a
execute()
不拆就直接干活
Phase 2b
synthesize()
子结果向上合并
Same task (write 4 independent Python modules), under a capped single-call budget: a single Agent scores just 25, while Sprout takes the full 100.
Scored automatically by a programmatic rubric, out of 100, reproducible. The single Agent runs out before finishing and gets truncated by token limits; after Sprout splits, each child node gets its own budget, and all 4 modules ship — the core value isn't parallel speedup, it's breaking past the token-budget bottleneck of a single call.
Two-phase Worker
analyze() uses a lightweight call to first judge “should this split?”, then execute() does the real work. Separating analysis from execution makes the split decision sharper.
Approach injection
When a parent splits, it generates a methodology and focus for each subtask and injects them into the child Agent's system prompt — roles emerge from the task rather than being predefined.
Safety bounds
max_depth · max_children · max_total_nodes · max_total_tokens — four ceilings to keep the tree from exploding; 24 unit tests cover the core modules.
Proven in real business.
Two internships, putting Agents into real business.
Sugon · Agent Development Intern
2025.12 – 2026.02Agent application development — Built an intelligent HR Agent and enterprise RAG Q&A on Dify, with vector-store integration and chunking-strategy tuning, reaching 85%+ Q&A accuracy.
Multi-agent cross-validation — Three specialized review Agents independently assess SFT training items, with structured scoring + conflict arbitration for joint decisions — cutting manual review cost significantly.
SFT data engineering and QC — An automated QC workflow on Feishu validates 500+ items a day, cutting manual effort from 3 hours to 10 minutes; reviewed Tool Calling and CoT correctness line by line.
Fanquan · AI Product Intern
2024.11 – 2025.02Product iteration and validation — Drove 4 versions of the AceEssay AI-reduction tool, with a dual Turnitin / GPTZero evaluation framework, bringing the AI-detection rate from 100% down to 10–20%.
Content growth and SEO — 60+ pieces of content drove 75K site visits and grew followers from 0 to nearly 30K; core-keyword ranking went from 48 to 9, with organic traffic up roughly 3× month over month.
More work.
ViT image classification
Fine-tuned ViT / DeiT with layer-wise learning rates + Label Smoothing + RandAug / Mixup / CutMix, reaching ≈98% validation accuracy.
PyTorch · TransformersReinforcement learningDQN multi-agent transport
4 Agents on a shared-network DQN with a collision-avoidance priority scheme — 95% round-trip transport success, average steps down 20%.
DQN · Multi-AgentEaster egg · the site itselfA personal website with a built-in AI twin
content/ is a single source of truth driving both the page rendering and the AI twin's system prompt — the prompt shown on the architecture page is the exact one running in production; real AI → demo mode → static page, a three-tier fallback, with the whole build log public.
Next.js 16 · AI SDK v6Easter egg · this very website
This keynote is itself an exhibit.
Read the full build log (in Chinese) ›Built with Claude Code, every step from one vague request to launch is on the record — the real prompts, the time each stage took, and every time the AI went off the rails and how I corrected it, all public in the build log (written in Chinese).
Tech specs.
One more thing.
This one is alive.
Every product demo above was an animation. This window is the real thing — my AI twin is live: it answers from a real résumé, and says so plainly when something's outside what it knows. How it works ›
Hi — I'm the AI twin of the person behind this site. I answer from their real résumé and project history, and I'll tell you straight when something isn't in my notes. Start with one of these:
Tell me about yourself
I'm Huang Yihang's AI twin. He's an AI master's student at Monash University (Class of 2027 — coursework ends 2026.07, degree conferred 2026.10), targeting AI application / Agent development roles, based in Chengdu and open to remote. He did an Agent development internship at Sugon; the intelligent operations system he built solo for Chenxi Flowers is live in production at a real company, with 5 Feishu group AI Agents handling order taking, outbound stock checks, cost logging and business questions; and PowerMyWeb (8 specialized Agents generating portfolio sites) is officially live with Free / Pro subscriptions open. He also independently built the open-source desktop AI assistant NoWorries, landed a merged PR on the open-source project OpenClaw, and wrote his own multi-agent framework, Sprout. Want the full picture? See the Projects and About pages.
When can you start?
He can start a remote internship right now; his master's coursework wraps in July 2026, and he can start full-time as early as August 2026 (the degree certificate arrives October 2026 — no blocker for a remote start). He's based in Chengdu and comfortable working remotely. In short: internship anytime, full-time from August 2026.
Are you open to remote work?
Yes—remote-friendly, and happy to come on-site or travel when it matters. He's based in Chengdu and open to both internship and full-time roles. The remote toolchain (Feishu, Git, async communication) is something he's actually run, both during his Sugon internship and while building this site collaboratively.
What sets you apart?
In one line: building from scratch, reading and refactoring someone else's system, and shipping Agents into real enterprise workflows—he has verifiable work in all three. From scratch: Sprout, a multi-agent framework of his own design. Refactoring: NoWorries, a desktop AI assistant deeply rebuilt on open-source Openwork (three-tier memory + safety sandbox), and OpenClaw, whose memory system he refactored—plus a separate fix PR merged upstream. Real business: at Sugon he landed multi-agent cross-validation inside a real SFT quality-control pipeline. Most candidates can show one of the three. He has the real thing in all three.
Pick a project and tell me about a hard bug
Here's a real one: while refactoring OpenClaw, he found the official token estimate for Chinese (CJK) was off by roughly 40% on the low side—which threw off every upstream context-trimming strategy and blew the token budget constantly in Chinese scenarios. He added a CJK-aware token-budget correction layer to stabilize it. More wrong turns and corrections (including the ones the AI itself made) are all laid out in the build log—ask me about any detail and I'll take it down to the mechanism level.
How do I reach you?
Email: 1653120857@qq.com, GitHub: github.com/hlbbbbbbb. You'll also find every contact method on the About page, or you can download the résumé PDF / save the digital business card. And if you're a recruiter—for anything that matters, email him directly; he replies fast :)
AI-generated answers can be off — for anything that matters, go by the résumé and a direct conversation with the real Huang Yihang. Chats may be logged to improve the twin.
Fallback chain: real AI → FAQ demo → static page · fully readable with no JS
Hiring for AI application / Agent development? Let's talk.
My master's coursework wraps in 2026.07 and I can start full-time as early as 2026.08 — looking for a role in AI application / Agent development. Send an email — I'll reply fast.
For recruiters · grab it in one click
