Transactions auto-classified
OpenAI structured-output layer categorized 10K+ transaction line items and de-duplicated the taxonomy at Studio PAV.
Open to SWE / AI Engineer roles · US
$
Xiang Hao
I build LLM-powered products and the async Python backends underneath them — multi-model routing, retrieval, RBAC'd APIs, and the pipelines that keep them honest. From audit-grade clinical data platforms to a multi-model AI assistant now in iOS and Android beta.
Client
React · TypeScript · Next.js · Capacitor
API
FastAPI async · OAuth2 / JWT · RBAC
AI layer
Multi-model routing · retrieval · structured output
Data
PostgreSQL · Supabase RLS · Pandas ETL
Runtime
Docker · Kubernetes · CI/CD · Prometheus
Impact
Numbers I actually moved
Each one maps to a shipped system — the case studies below explain how.
Cold-start latency
Mo Paws first-response latency cut 75–80% through knowledge-base preloading and cost/latency-aware model routing.
Audit-cycle time
Async FastAPI services with PostgreSQL, hierarchical storage, RBAC and OAuth2 on Azure at EDA Clinical.
Manual work removed
Teams transcripts → structured summaries → Trello cards and checklists, org-wide, with no human in the loop.
Selected work
Systems, not screenshots
Ten projects across AI, backend and data. The two marked Case study go deep on the architecture and the trade-offs.
Case study
Mo Paws — multi-model AI pet companion
A cross-platform pet-care assistant built at Rocha Tech LLC, now in closed TestFlight beta: breed-aware answers, behaviour assessment, health scheduling. I own the whole stack — auth, data model, retrieval, and the model-routing layer that keeps responses fast and cheap.
206 breed profiles + 181 knowledge facts (205/205 of the stored AKC baseline) · cold-start latency −75–80% · cost/latency-aware routing across DeepSeek, GPT-4.1-mini and Claude Haiku with provider fallback.
Case study
Time Tracker — role-based timesheet platform
An internal platform for EDA Clinical replacing a spreadsheet-and-email approval chain. Three role-scoped interfaces (developer, manager, admin) on one async FastAPI service.
OAuth2 + JWT with RBAC enforced at the dependency layer · configurable approval routing · audit logging on every state change · deployed on Azure with Docker + CI/CD.
Job Radar — job-intelligence agent
A daily agent that pulls postings, scores fit against a profile, and emits a structured JSON feed with match tier, category, sponsorship signal and application status.
LLM scoring behind a strict JSON schema with validation + retry, so downstream tooling never parses free text.
Industrial Safety Detection
A ResNet-50 vision pipeline that flags PPE and equipment-safety violations on factory and construction footage, streamed through Kafka and served from AWS ECS.
Replaced 30+ hours/day of manual review; blue-green Docker deploys kept the review queue live during releases.
Dataset-JSON Viewer
A browser tool for exploring CDISC Dataset-JSON clinical datasets — nested-structure navigation, column filtering, and a preview that stays responsive on large exports.
Built for reviewers who previously opened multi-hundred-MB exports in Excel.
House Price Prediction
A model bake-off on the Ames housing set: linear regression, kernel perceptron, decision trees, FNN, 1D-CNN and 2D-CNN, compared on identical feature pipelines.
Graduate ML coursework at UC Davis — the interesting part is the ablation across model families, not the leaderboard score.
Global Development Insights
A linked-view D3 dashboard — parallel coordinates plus bar and pie views that cross-filter one another. The coordinates plot 154 countries across 20 development indicators.
Hand-built brushing and linking — no charting library doing the interaction for me.
COVID-19 Rate Tracker
A desktop tool that pulled regional infection data on a schedule, cached it locally and rendered comparative rate curves with a searchable region picker.
Persistent local cache so the UI stayed usable when the upstream feed went down.
Karting Game
A 3D kart racer in Unity with three environments, custom drift physics, lap timing and an AI opponent that follows a waypoint racing line.
Built to learn real-time systems: fixed-timestep physics, object pooling, frame budget.
Cold Call System
A classroom tool that samples students without replacement and shows an "on-deck" queue so the next person can prepare instead of being ambushed.
Weighted sampling with a participation ledger persisted between sessions.
Capability matrix
What I reach for, and where I've used it
Levels are self-assessed against production use — the right-hand label says which system it came from.
AI & LLM engineering
primary trackLLM API integrationOpenAI · Claude · DeepSeek
Structured output + schema validationStudio PAV · Job Radar
Retrieval: chunking · ranking · hybridEDA digest · Mo Paws
Multi-model routing & fallbackMo Paws
Prompt regression testinggrowing area
Backend & APIs
primary trackPython5 yrs · daily driver
FastAPI (async)EDA Clinical · Yunlv
Auth: OAuth2 · JWT · RBACTime Tracker · Supabase RLS
Queues, idempotency, async workflowsTemporal · task pipelines
Javarobotics scheduling
Data & ML
supportingPostgreSQL: schema + tuningEDA · Supabase
Pandas / ETL pipelinesSilentNight · Studio PAV
CDISC SDTM / ADaMclinical-trial data
PyTorch / computer visionHRG · robotics
D3.js / data visualisationEDA digest · dashboards
Cloud & delivery
supportingDockerevery project
CI/CD (GitHub Actions)5+ microservices
AzureEDA Clinical production
Kubernetesrollouts, health checks
AWSCertified Cloud Practitioner
Experience
Where the work happened
Software Engineer
Rocha Tech LLC May 2026 — Present · Remote- Building Mo Paws, a cross-platform AI pet-care assistant now in closed TestFlight beta — owning auth, data model, retrieval and model routing end to end.
- Designed a task-based multi-model router across 3 providers with cost- and latency-aware selection and automatic failover; cold-start latency down 75–80% through knowledge-base preloading and SSE streaming.
- Built deterministic retrieval over a curated knowledge base — 206 breed profiles covering 205/205 of the stored AKC baseline, plus 181 canine knowledge facts — with image-based document extraction and safety guardrails.
- Enforced multi-tenant isolation in PostgreSQL Row Level Security policies rather than application middleware, shipping schema and policy together in versioned migrations.
Data & AI Engineer
Studio PAV Feb 2026 — May 2026 · Contract, Remote- Designed an end-to-end Python/Pandas workflow merging Shopify and CRM records across 10 years, unifying tens of thousands of transactions into analysis-ready datasets.
- Built a feature module computing 20+ behavioural features per customer, powering percentile-based segmentation tiers exported as multi-sheet Excel.
- Shipped an OpenAI structured-output classification layer that categorized 10K+ line items at 90% accuracy and collapsed a duplicated taxonomy.
Software Engineer
EDA Clinical Oct 2024 — Dec 2025 · Pennsylvania, US- Automated Teams transcript → Trello documentation with Microsoft Graph and OpenAI — structured summaries, cards and checklists — removing 5+ hours/week of manual meeting write-ups org-wide.
- Built a React + D3 RAG-powered weekly digest with semantic retrieval, lifting stakeholder engagement 20%.
- Developed async FastAPI services with PostgreSQL, hierarchical storage, RBAC and OAuth2 on Azure — audit-cycle time −25%.
- Standardized health checks and CI/CD across 5+ Dockerized Kubernetes microservices, cutting rollout time 30%.
- Built the SilentNight ETL pipeline for clinical-trial lifecycles with 100% CDISC SDTM/ADaM compliance, and integrated the CDISC Open Rules Engine for end-to-end audit traceability.
Software Engineer Intern
HRG Institute of Research & Innovation Jul 2021 — Sep 2021 · Anhui, CNAutomated safety review with a PyTorch ResNet-50 computer-vision pipeline fed by Kafka and served from AWS S3 + ECS. Dockerized blue-green deployments removed 30+ hours/day of manual review.
Software Engineer Intern
Shanxi Yunlv Tianxia Network Technology Jan 2021 — Mar 2021 · Shanxi, CNAutomated expense approval and ticketing with FastAPI and Temporal workflows, then added Prometheus, Grafana and Jaeger observability for a rollout serving 1,200+ employees across six departments.
Software Engineer Intern
Shanxi Anxing Hengchuang Robot Tech Jul 2020 — Sep 2020 · Shanxi, CNImproved warehouse-robot navigation by pairing laser sensors with a reinforcement-learning course-correction policy in PyTorch, added OpenCV + SQL delivery-mismatch detection, and streamed battery telemetry over Kafka to schedule charging priority in Java.
About
Short version
I'm a software engineer with an MSCS from UC Davis (2024) and a BS from the University of Oregon. Most of my work sits where an LLM meets a real system: the model is the easy part, and the retrieval, schema, permissions and failure paths around it are the job.
At EDA Clinical I spent a year in regulated clinical-data land — audit trails, CDISC compliance, RBAC — which is a good place to learn that "it works on my machine" is not a deliverable. Since then I've been at Rocha Tech LLC building Mo Paws, a multi-model AI pet companion now in closed TestFlight beta — end to end: Supabase auth and RLS, the retrieval layer, the model router, the mobile shell.
I'm looking for AI Engineer / Applied AI / Backend roles in the US where the product actually ships to users. Fastest way to reach me is email.
Based inWilmington, DE · US
EducationMSCS, UC Davis
FocusApplied AI · Backend
LanguagesEnglish · 中文
Education & certification
M.S. Computer Science
University of California, Davis
2022 — 2024
B.S. Computer & Information Science
University of Oregon
2019 — 2022
AWS Certified Cloud Practitioner
Amazon Web Services
2025
Contact
Let's talk
Hiring, contracting, or just want to compare notes on LLM plumbing — either box works.
Emailxianghao98520@gmail.com
LinkedInin/xianghaosuc
GitHubXiang-Suc
LocationWilmington, DE · open to relocate