Yupo (Jason) Niu
Eight years at an HKEX-listed company on pricing, credit risk, and approval engines — now rebuilt with LLMs, RAG, and agents. I pick up unfamiliar stacks fast, own delivery end to end, and turn unclear requirements into shipped systems. Claude Code, MCP, and custom tools are how I work faster now.
A repeatable method for unfamiliar problems, refined over 17 years.
AI Engineer · Calgary, Canada
Who I am: I'm Jason — 17 years building production systems, including 8 years designing decision engines, risk models, and workflow automation at Edianyun (HKEX-listed). Before LLMs, I built rule-based systems that automated pricing, risk, and approvals. Now I build the same class of systems with LLMs, RAG, and agents. 3.96 GPA across Data Analytics (4.0) and Integrated AI (3.92) at SAIT.
What drives me: I'm curious about how things work — and how they could work better. I don't use AI just to code faster. I use it to learn faster: custom curricula, unfamiliar stacks, new domains in days.
// what 17 years actually built
I ship in stacks I've never written.
Rewrote Edianyun's early online mall from PHP to Java with no prior PHP — delivered on schedule when PHP talent was unhirable and the stack couldn't scale horizontally. Since then: self-taught React, Go, and Swift with AI, shipping native apps and CLI tools.
Eight years. From first spec to production.
As Team Lead at Edianyun I owned the projects I led — pricing, risk control, the rebuilt permission platform — from requirements and analysis through architecture, tech specs, delivery tracking, deployment, and code review.
0→1 where there was no playbook.
Joined Edianyun at the startup stage and turned undefined business — leasing price calculators, multi-bureau credit risk — into shipping code by working hands-on with finance and the CFO. Brought in containerization and microservices to make ops and CI/CD safer.
The methodology runs on a toolchain I built myself — voice input, a key vault, task notifications.
Dictate code notes in English or Chinese
Claude Code drives edits and review loops
Typed, tested, reviewed
CI/CD to production in minutes
Tools I built for voice-driven development
Voice-to-clipboard pipeline: Groq Whisper ASR → Gemini AI correction → instant IDE paste. Supports bilingual Chinese-English with programming terminology.
Bash · Groq · Gemini
AES-256-GCM encrypted API key manager built in Go. Secure credential storage via macOS Keychain + SQLite for multi-service development.
Go · AES-256 · Keychain
Real-time Telegram notifications for Claude Code task completion. LLM-powered session summarization via Groq for async monitoring.
Bash · Telegram · Groq
84
Active Repos
past year
1,139+
Commits
past year
25
Projects
in development
9
Languages
in production
Every role began with a system that didn't exist yet. Here's what I shipped.
Help build a smart building management system — BMS, mobile app, AI document search, and face-recognition access control — for a Calgary startup (company dissolved Apr 2026)
Shipped the RISE mobile app to the client (App Store + Google Play) and prototyped the AI RAG workflow. The broader BMS was not delivered — the company dissolved in Apr 2026.
Build and scale enterprise decision systems for a publicly traded company — financial algorithm engines, real-time risk control, workflow automation, and distributed architectures handling 10K+ concurrent users across 8 product platforms
Led 3 teams across 8 product platforms — financial algorithm engines, risk control, and workflow automation — in production at an HKEX-listed company.
Build financial systems for a leading crypto-mining hardware company — mining pool revenue distribution, a Bitcoin payment app, and automated trading
Independently shipped the internal Bitcoin payment system. Early hands-on with distributed computing and real-time data pipelines.
Deepen AI specialization — computer vision, predictive modeling, AI governance, and applied AI project delivery
Completed — 3.92 GPA. Combined with Data Analytics, total 16 courses, 15/16 A+, overall 3.96 GPA across 54 credits
Build data foundations — statistical analysis, predictive analytics, and business intelligence to complement 17 years of engineering experience
4.0 GPA, capstone ML system (97.4% accuracy) — statistics and ML foundation for the AI coursework that followed
Production and enterprise work first; learning experiments below. Filter by language and type — each card opens with the question the project set out to answer.
// matches: 24 / 24▌
Can an AI agent find, score, build resumes for, and apply to jobs autonomously?
Autonomous multi-agent job search pipeline orchestrating 5 LLM providers with circuit breakers, structured output, reflexive quality gates, and auto-apply via real Chrome. 1,250+ tests, 4 discovery sources, Telegram bot with inline keyboards.
How do you make enterprise documents searchable without structured metadata?
AI workflow system with intent recognition, routing, and multi-agent orchestration. Hybrid retrieval (BM25 + vector search) with pgvector, OCR document processing, and code execution.
How do you sell limited stock to 10,000 buyers at once and never oversell by a single unit?
Enterprise flash-sale and timed-auction platform (Edianyun) for liquidating retired office IT — the resale channel for a HKEX-listed IT-services firm managing 1.37M+ devices across ~51K enterprise customers. Built to survive 10K+ concurrent buyers with zero overselling: Redis atomic stock deduction via Lua, layered defense-in-depth (rate-limit → cache → atomic decrement → MQ), RocketMQ-async order pipeline, WebSocket real-time bidding, and a self-healing compensation loop that reconciles stuck inventory on failure.
How do you decide a customer's credit limit in real time without driving up bad debt?
Core credit risk system (Edianyun) deciding how much credit and deposit waiver each customer gets across the full pre-lease → in-lease → post-lease lifecycle. Multi-bureau credit scoring (A/B/C/F/L/O scorecards), 4-tier approval (M1→M2→M3→city manager), configurable credit-limit pools, and overdue collection — built on a multi-module Java microservice stack with MaxCompute data-warehouse integration.
How do you keep permissions accurate and low-latency when 1,000 people change roles constantly?
Company-wide authorization middleware (Edianyun) serving ~1,000 employees across a dozen internal systems with menu/button/URI 3-tier RBAC. Decoupled HR org structure from permission config via an auth-post mapping layer; Redis-first auth hot path with binlog CDC (Aliyun DTS) → RocketMQ cache sync, a lightweight HMAC-signed interceptor SDK, and JavaParser-based endpoint auto-scanning.
How do you turn six independent business databases into one audit-grade financial ledger — consistently, idempotently, and without blowing memory on tens of millions of rows?
Pure-Java financial ETL service (Edianyun) built for HKEX IPO audit compliance — it reconciles a leasing business's receivable / received / prepaid / hedging and aging ledgers across six upstream systems (Oracle EBS, Hyperion HFM, U8C, order-billing, risk-control credit, ERP asset core) and reports them into Oracle EBS / Hyperion in the exact format Finance and the auditors require. Divide-and-conquer extraction: one wide read pulls every needed field, large pull jobs of millions–tens of millions of rows split into small tasks running in parallel on a Fork/Join pool behind a join barrier, results staged in Redis (long TTL + async eviction) to avoid re-hammering the read replicas. RocketMQ choreography gates each transfer behind an Oracle master-table state machine — request_id idempotency plus phase/status polling — while dynamic-datasource @DS isolates every source's connection pool, achieving cross-system consistency without a distributed lock.
How do you let an LLM agent use your existing services without handing it the keys to everything?
Wraps a traditional Go ticketing backend with the Model Context Protocol so LLM agents (Claude Code, Codex, Gemini CLI) can book, pay, refund, and validate demo tickets through MCP tools, resources, and prompts — behind a Figma-style human consent + scope gate. The bridge stays thin; correctness and authorization live in the REST/infra layer: scoped expiring tokens, identity derived from the token (never from tool args), and concurrency-safe booking that prevents double-sell and double-charge.
How do you keep four microservices consistent when any step can fail, retry, or be delivered twice?
End-to-end demo of distributed-transaction correctness: a ticket-purchase Saga across four Spring Boot microservices orchestrated by Temporal, with a transactional outbox → RabbitMQ → live SSE pipeline, idempotent steps behind a barrier table, and automatic compensation in reverse on failure. Visualized live in a Next.js + React Flow board.
Offline Python analytics platform (Edianyun) — the data-science counterpart to the online Credit Risk Control engine. It builds a customer-level wide table from the leasing business's MySQL + MongoDB stores and 16 credit bureaus, engineers WOE/IV features, and trains interpretable logistic-regression scorecards validated with KS/AUC/PSI. Its output is weight-optimization advice written back to the online credit engine — turning hand-tuned scorecard weights into data-driven ones learned from human-review decisions and post-rent bad-debt outcomes. Around that interpretable core, the modeling toolkit extends to GBDT challengers, anomaly/graph anti-fraud, survival-style time-to-default analysis, and reject-inference validation — model choice driven by the data's characteristics (mixed types, 60–70% missing financials, 1–5% fraud, right-censored labels, graph structure), not by chasing the newest algorithm. Read-only replicas and an isolated analysis DB keep modeling fully decoupled from live authorization. Reconstructed from the real codebase.
LLM quality automation system using LLM-as-Judge pattern. Local Qwen 3.5 simulates e-commerce customer service, DeepSeek V3 evaluates responses on relevance, faithfulness, and effectiveness — achieving 0.94/1.00 composite score.
At Havenz Tech (Calgary, dissolved Apr 2026): solely built and shipped the RISE React Native mobile app (iOS + Android, published to the App Store and Google Play) for a Calgary sports-venue membership system; prototyped a LangChain document-RAG workflow (BM25 + pgvector hybrid retrieval, PaddleOCR ingestion, validated in team testing); and researched face-recognition hardware for a regulated building-access requirement (R&D, not productionized). CI/CD on GitHub Actions + Docker + GCP Cloud Run. The broader building-management system was not delivered.
Multi-user SaaS trading system where local LLMs (Qwen 2.5, DeepSeek-R1) make entry/exit decisions. TradingView webhooks trigger signals, AI evaluates market context via customizable prompts, and trades execute through IBKR API with per-user/per-symbol configuration.
Lightweight daemon-mode API key vault for AI platforms. Secures keys with AES-256-GCM encryption and macOS Keychain integration, serves them to local scripts via Unix socket API. Supports 7 AI providers (OpenAI, Anthropic, Gemini, Groq, DeepSeek, Qwen, GLM).
Dual-endpoint (CLI + iPhone Web) voice-to-text tool with a full ASR → AI correction → pronoun verification pipeline. Uses Groq Whisper for transcription with Gemini 2.5 Flash for intelligent text correction, served via Cloudflare Tunnel.
Native macOS voice-to-text dictation tool with global hotkey activation. Supports 8 ASR services (OpenAI Whisper, Groq, Alibaba, Tencent, Baidu, iFlytek, AssemblyAI, Speechmatics) with cost optimization and multi-language localization.
iOS app for real-time golf swing biomechanical analysis using Apple Vision framework. Detects 7 biomechanical metrics and 11 issue types at 30 FPS with pose estimation, providing scoring (0-100) and video import support.
Customized BiliBili streaming client for Apple TV (tvOS): QR login, live streaming with real-time comments, HDR/Dolby Vision support, SponsorBlock integration, and playlist management.
Led the migration of legacy Tomcat applications onto a company-standard Spring Boot scaffold running on Alibaba Cloud ACK (managed Kubernetes), as DevOps team lead working hand-in-hand with ops. Authored the reusable Jenkinsfile / Dockerfile / docker-compose templates and Jenkins pipeline configs, and stood up isolated dev / test / stage environments for large company-wide projects. Built the observability layer on ELK + Skywalking with sidecar injection — designed Kibana dashboards and log-based alerts across multiple services and drove periodic system optimization from the findings.
The hands-on operational backbone behind 8 years of Edianyun's Java platform. From 2016 there was no separate ops team — I ran the Alibaba Cloud CentOS ECS/VPS fleet myself: daily SSH troubleshooting, a large body of shell scripts for deploys, log triage, and data import/export, routine JVM tuning under load, and Nginx setups (TLS certs, reverse proxy, location rules) for front- and back-end test environments. As the platform grew I automated repeatable ops with Ansible and drove container workflows on kubectl / Minikube / Lens. Self-taught Terraform on throwaway VPS instances to practice infrastructure-as-code. macOS + Vim power user since 2012.
I design custom AI curricula to ramp into fast-moving tools.
GitHub trending, AI research on X, and custom learning paths to get working examples running in unfamiliar areas. Current threads:
From LLM orchestration to distributed backends — 17 years shipping production systems.
42 skills·7 domains·17 years
Agentic orchestration, RAG, LLM reliability
Models, computer vision, data pipelines
Java, Python, Go, TypeScript, Swift
Distributed systems, queues, databases
Linux, containers, CI/CD, IaC
Claude Code, MCP, local AI tooling
English + Mandarin
I'm looking for AI, backend, or platform teams in Canada where reliability and judgment matter. Let's talk.