Pengcheng Xu, builds intelligent systems.
I design and ship AI-powered applications and real-world products — where systems thinking, mobile craft, and language models meet.
Three things I build — with receipts.
Each of the next four featured projects lives in one of these buckets. The numbers are what shipped, not what was promised.
Hybrid retrieval, citation-grounded LLMs, multi-dimensional scoring. Designed for recall, latency, and refusal — not demo-day magic.
AI systems where retrieval beats fine-tuning
On-device vision, AR navigation, real-time guidance. Apple's frameworks pushed past the obvious — interface as instinct.
iOS where intelligence feels native
Smart contracts, async ingestion, cache strategy, reliability budgets. Boring infrastructure — measured, monitored, done well.
Backend that holds up under traffic
A simple loop. Repeated obsessively.
Same three steps every time. Below each, a moment from a real project where the step paid off.
Understand the problem
Deeply. Before code, before architecture — what are we actually trying to make easier? Who's on the other side of the screen?
Design scalable & intelligent solutions
Pick the right primitives — embeddings, queues, caches, agents. Sketch the system before the implementation.
Optimize performance & UX
Measure what matters: latency, recall, retention. Then make it feel inevitable to the person using it.
DevPulse — making 50,000 documents
feel like one searchable mind.
Engineering teams write more than they can read. Important context lives buried in docs, postmortems, code comments, Slack threads. Search is keyword-shallow; LLMs hallucinate without grounding.
A retrieval-augmented platform that combines dense embeddings, hybrid keyword ranking, and a small reasoning layer — so questions return the source, not a guess.
baseline · 50K-doc corpus
scaling independently
- +23% Recall@5 vs. pure-dense baseline
- Multi-service architecture lets the AI worker scale without touching the API tier
- Redis Pub/Sub eliminates polling, keeps streaming responses snappy
Hybrid retrieval (dense + sparse via RRF) lifted Recall@5 by 23% on a 50K-document corpus. Java Spring Boot for the API tier, Python FastAPI for the AI worker — each scales without touching the other. Redis Pub/Sub bridges them at streaming latencies.
GreenTraceChain — turning every gram of CO₂ into auditable, on-chain truth.
Carbon accounting and offset markets are riddled with opaque data: emissions are self-reported, ledgers can be quietly amended, and regulators see numbers days — sometimes months — after they matter. The "double-carbon" goal needs trust the current system can't provide.
A four-role carbon platform on ChainMaker (长安链): enterprises file emissions, auditors review & sign, regulators watch the ledger live, and admins mint carbon credits — all stitched together by five smart contracts that make every record immutable, traceable, and quantum-safe.
500+ TPS sustained
via refined block indexing
- Led a 5-member team architecting the chain + service layer
- Industry-specific carbon models (2 verticals shipped) for accurate accounting
- CarbonCoin rewards convert verified reductions into tradable credits
- Credit-scoring + clearing keeps the market honest under partial trust
- Won First Prize at the 2024 China Service Outsourcing Innovation Competition (Aug 2024 · A21 · 云象网络)
Led a 5-member team architecting the chain + service layer. Won First Prize at the 2024 China Service Outsourcing Innovation Competition (题号 A21 · 云象网络) with team 你的外包我来包.
DeepNews-LLM — scoring news on eight dimensions
before you trust it.
// internal · 03 · factual_accuracy retrieve(article, k=8) → evidence extract(article) → claims[] score = max(0, 10 − 0.5 × |conflicts|) refuse if confidence < 0.6
Generative AI made fabricating a believable headline a one-prompt task. Single-signal fact-checkers — source blacklists, keyword filters — miss the patterns that make a story feel true while being false. Readers need a second opinion that thinks, not just blocks.
A LangChain-based agent that scrapes the article, then scores it on five internal dimensions (relevance, logic, facts, tone, causality) plus three external ones (source, prior debunks, cross-source corroboration). AHP × entropy-weighted aggregation gives the final verdict — streamed back over SSE as the agent thinks.
over labeled benchmark
via async streaming
- POST /analyzeWeb · full 8-dimension report from a URL
- POST /chat · conversational follow-ups with persisted session history
- POST /chat/resultQuestion · ask why a specific dimension scored what it did
- Microservice-shaped: gateway · agent · scoring · extraction split for independent scaling
Focuspace — a camera that sees with you,
not just for you.
Move until it feels right.
No language, no menus — just a breathing dot that tells you where to point. The simpler the cue, the more the photo feels like yours.
People who travel with their iPad or phone don't take bad photos because their camera is bad — they take bad photos because nobody ever taught them composition. The viewfinder is silent. Symmetry, rule of thirds, leading lines… you either know them, or you don't.
An iPad camera app submitted to Swift Student Challenge 2026. Pick one of nine composition templates; a breathing white diamond floats around the viewfinder, telling you exactly which way to move. Move until it feels right — no jargon, no menus.
vs. native camera UX
−40% overlay jitter under motion
- "Move until it feels right" — no language, just a dot you chase
- Submitted to Apple's Swift Student Challenge 2026
- Settings expose Preview Grid · Guidance UI (Scope/Missing/Arrow) · Template Overlay · Foundation Models toggle · Debug HUDs
- Ships as a single Swift Playground (.swiftpm) — Xcode or iPad direct
A quick scan through the rest.
Side projects, course capstones, and weekend experiments where I tested ideas before they became real products.
ClaudeMenuBar
A macOS menu-bar companion for Claude Code. NWListener HTTP hook + Combine FSM cuts app-switching by ~55% and AI-interaction latency by ~65%.
RuleERU
An EU compliance platform for cross-border e-commerce — automated tax declarations, multilingual support, regulation knowledge base.
MemoTree
An AI + AR tourism app — ChatGPT-powered route planning (−30% trip-planning time) plus ARKit landmark recognition at 98.5% precision.
FitKid
A childhood-obesity intervention app — Create ML food recognition at 92% accuracy, AR exercise tracking that lifted training adherence by 35%.
I don't just use build with AI
— I build systems with AI as a primitive.
Production RAG & LLM pipelines
Embeddings, hybrid search, reranking, citation-grounded generation. Designing for recall, latency, and refusal — not demo-day magic.
Multi-agent coding workflows
Planner / executor / verifier loops that turn vague intent into reviewed pull requests. Tooling, sandboxing, observability.
AI-assisted development
Heavy hands-on with Claude Code, Codex, and Cursor. I treat the model as a collaborator with a code review checklist.
Real systems. Real users. Real numbers.
The work below shipped to production and was measured by people I had to keep happy.
Owned backend modules across the company's commodity-trading platform — Material Inquiry System, Product Library CRUD, and the price-trend visualization API. Took features from spec to production, then iterated on latency and reliability under real traffic.
-
10K+
Active users on the Material Inquiry System I shipped.Sustained through full-traffic ramp without incident.
-
−40%
Query latency reduction on the hottest path.Cache strategy + query plan rework + N+1 elimination on Java/Spring Boot + MySQL.
-
+30%
Reliability gain on the Product Library CRUD & REST surface.Idempotent writes, retry-safe APIs, defensive validation at the boundary.
-
−20%
Latency cut on the price-trend visualization module & its API.Faster UI response — let traders make decisions in real time.
Three lists, kept short on purpose.
I can hand you a spreadsheet of acronyms. I'd rather show you what I reach for first.
Things I think in.
- Java · Swift · Python · Go
- C++ · JavaScript / TypeScript
- SQL
Things I build with.
- Spring Boot · MyBatis · Gin · Gorm
- SwiftUI · AVFoundation · ARKit
- React · Vue · Node.js
- FastAPI · LangChain
Where intelligence lives.
- PostgreSQL · MySQL · MongoDB · Redis
- Elasticsearch · pgvector
- Core ML · Create ML · FoundationModels
- PyTorch · TensorFlow · Hugging Face
Boring infra, done well.
- Linux · Git · Docker · Kubernetes
- Kafka · REST · Nginx
- Postman · Dialogflow
- Claude Code · Codex
A few shiny things.
National prizes, regional firsts, an Apple certification, and a top-1% scholarship — receipts for the work above.
-
Aug 2024
First Prize · National
2024 China College Service Outsourcing Innovation & Entrepreneurship Competition — National First Prize for GreenTraceChain.
National · Team lead -
2024
National Innovation Project
National College Student Innovation & Entrepreneurship Training Program — project approved at the national level under China's MOE.
National · PI -
Nov 2024
National Scholarship
China National Scholarship — awarded to the top 1% of students at the university level.
Top 1% · University -
2025
First Prize · East China
2025 Mobile Application Innovation Competition — regional First Prize in the East China division.
Regional · iOS -
Jul 2023
Second Prize · East China
2023 Mobile Application Innovation Competition (East China division) — for an early iOS prototype that seeded MemoTree.
Regional · iOS -
2023
Third Prize · Zhejiang
2023 Zhejiang Provincial College Service Outsourcing Innovation Application Competition — provincial Third Prize.
Provincial · Team -
Sept 2023
Apple Certified
App Development with Swift Associate — Apple's international certification for foundational iOS app development.
International cert