Founding Engineer, AI Data & Agent Platforms

YINFEI WANG

646-***-**** ****************@*****.*** github.com/Steven-wyf linkedin.com/in/ywsteven/ EDUCATION

New York University, New York, NY GPA: 3.8 M.S. Computer Engineering, Tandon School of Engineering

Sep. 2024 – May. 2026

University of California, Irvine, CA B.S. Computer Science and Engineering Sep. 2016 – Mar. 2021 PROFESSIONAL EXPERIENCE

Topify.ai San Francisco, CA

Founding Engineer Nov. 2025 – Present

GEO analytics platform — 154 merged PRs / +200K LOC across 4 codebases (Rust/Rocket, Next.js 14 App Router, Deno + Oak agent, Python/FastAPI admin); 61 ADRs / design specs / production runbooks sustaining spec-first development at one doc per 1.9 days.

• Project Lead — AI Brand Visibility Pipeline & AI Heatmap

Primary maintainer of daily LLM data-fetch pipeline (Apalis V2 distributed task queue on PostgreSQL + Supabase, RLS multi-tenancy, 5-provider fan-out with async 3-step submit poll retrieve, PgBouncer connection pool budgeting 167/170 ceiling) — pipeline runtime 8h 4h (50%).

Deployed sentiment worker on AWS ECS Fargate (XLM-RoBERTa via ONNX Runtime in Rust) with AWS Lambda + EventBridge + CloudWatch P95 + SNS alerting; owned reliability layer — lock-free circuit breaker, bounded- concurrency recovery semaphore, DataForSEO rate-limit defense.

Shipped AI Heatmap across 3 codebases — 8 backend endpoints (REST/OpenAPI), 34 React 18 components (shadcn/ui + Radix + Tailwind + Zustand + Recharts + TipTap), 24 typed Next.js 14 App Router BFF routes + TanStack Query v5 + centralized error code registry.

Delivered first practical Share of Voice (SOV) metric for generative search — 71-file coordinated single-day big-bang release; introduced repo's first Playwright E2E suite.

• Project Lead — AI Agent Platform — proposed and led migration from LangChain + LangGraph Deno

+ Zypher (CoreSpeed open-source on JSR), 4 decision criteria; owned 5 subsystems = +19K net LOC

(22%)

Architected GEO Agent on Deno + Oak + Zypher with three-way routing decision framework, supervisor loop + mode switching (2 infra savings), 17 self-registering agent skills enabling multi-step agentic workflows.

Integrated 5 MCP (Model Context Protocol) servers across 3 transport modes (replacing LangChain @tool decorator) + output secret scanner blocking LLM credential leaks; reliability middleware (retry taxonomy + AbortController cancellation propagation).

Built Helper Bot floating chat with page-context injection via React Context (RAG / in-context adaptation) — answers with on-screen data not generic advice; bilingual EN/ZH doc router + 35 contextual <HelpLink> entry points.

Designed 7-phase Heatmap topic-generation pipeline in Deno: site profile bootstrap LLM clustering per-topic prompts regenerate session view finalize with A/B/C1/C2 classification.

• Project Lead — Production Reliability (3 P0 post-mortems) & Cross-Stack Verification

Resolved PgBouncer pool exhaustion (1,999 duplicate jobs / 6-replica fleet) via advisory-lock leader election (transaction- scoped) + GitHub Actions CI/CD deploy-window guard, flagged by Macroscope static-analysis bot (32 review cycles).

ECR Lifecycle Bomb 5-phase 1-day response: emergency restore + immutable image tags + templated task-def + 25h CloudWatch silence alarm + formal ADR — pre-empting customer-facing outage masked by Fargate host cache.

Built user-to-user referral program on append-only two-ledger (event sourcing + append-only ledger accounting), 7-day refund grace, idempotent webhooks, collision-safe SQL CTE, 12 ADRs + refund SOP runbook.

Designed cross-stack algorithm mirror — Python admin independently re-implements Rust brand-matching algorithm, converting deliberate duplication into verification architecture catching multiple production drift events.

• Project Lead — Subscription State Machine & Polymorphic Payment Provider Abstraction

Owned subscription state machine for all 10 Stripe webhook event types in Rust + SeaORM migrations (TDD: 31 unit + 30 sweep tests, race-safe row-locking, idempotent webhooks) with daily reconcile cron.

Migrated Pricing V2 3 4 tiers with 19 ADRs + 13-branch derivePlanState pure-function single source of truth; designed 3-tier Stripe key permission isolation (read-only / payment-links / promo-codes) blocking privilege escalation.

Designed polymorphic payment-provider abstraction (3 traits) for Stripe + PayPal API + Alipay with zero behavior change; pre-implementation 5-agent design review caught $19-vs-$14 fraud-class billing bug before any code. Huawei Cloud Computing Technology Co. Ltd. Beijing, China Sr. Software Development Engineer Jul. 2021 – Aug. 2024

• Project Lead — Cloud Scalable NoSQL Database Design, Optimization & E2E Delivery Dec. 2023 – Aug. 2024

Led design and optimization of distributed NoSQL on MongoDB + RocksDB, implementing C++ routing layer enhancements + LSM-tree-based aggregation — 100k+ doc/sec throughput, ms-level latency under high concurrency.

Developed atomic bulk transaction mechanisms across gateway, forwarding, and storage layers (Kafka async cross-layer event propagation); standardized REST + SDK interfaces (multiplayer trading, large-scale payroll).

Built DevOps-driven CI/CD pipeline on Kubernetes + Jenkins + Docker with Python regression tests + Prometheus health monitoring + Shell backup automation — daily incremental releases, +60% deployment velocity.

• Project Lead — Intelligent Performance Monitoring & Optimization for KVS Service Jan. – Aug. 2023

Designed intelligent performance monitoring platform for KVS service, embedding FTDS metric tracking into C++ core modules to visualize latency/throughput/concurrency across MongoDB–RocksDB data layers.

Built custom metric visualization + benchmarking framework on Python + Prometheus + Grafana supporting user-defined parameters (KV size, concurrency, partition count) — +95% performance analysis efficiency.

Optimized high-concurrency I/O + request aggregation through RocksDB batch interfaces + caching — P99 latency −90%, throughput +86% under production workloads.

• Project Lead — Disaster Recovery Tool from System Log Data May. – Oct. 2022

Designed disaster recovery framework for distributed storage systems with REPL + GFlag-based interactive interface to extract, decode, and back up system logs after service crashes.

Implemented log decoder + recovery engine in C++ to parse persistent WAL logs, locate data corruption at bit level, reconstruct LSM-tree states ensuring full consistency to point of failure.

Authored user documentation + recovery playbooks for multiple disaster scenarios — streamlined developer onboarding, improved recovery success rate in system simulations.

• Project Lead — AWS-S3 Distributed File System Features Jul. 2021 – Jan. 2022

Delivered core features for distributed file system on AWS S3 (auto file recycling, duplication detection, system snapshots), enhancing storage efficiency + data reliability.

Developed TCC-based (Try-Confirm-Cancel) deduplication framework in C++ and Python with client-worker task queue for real-time cleanup of redundant files across distributed nodes.

Designed configurable snapshot mechanism with exponential-backoff retry logic ensuring consistency + fault-tolerant recovery in large-scale environments.

TECHNICAL SKILLS

Languages: Python, Rust, TypeScript, C++, JavaScript, SQL, Go, Java, C, SHELL Backend: Rocket, SeaORM, Apalis V2, FastAPI, Next.js 14 App Router (BFF), React 18, Deno + Oak, REST/OpenAPI, SSE, Node.js, Django, Spring Boot

LLM / AI: MCP (Model Context Protocol), multi-step agentic workflows, RAG / in-context adaptation, ONNX Runtime, XLM-RoBERTa, PyTorch, LangChain, LangGraph, Zypher, MLflow Database: PostgreSQL + Supabase, PgBouncer, RLS multi-tenancy, SeaORM migrations, MongoDB, RocksDB, LSM-tree, DynamoDB, MySQL, Redis, Amazon S3

Cloud / Infra: AWS ECS Fargate, S3, Docker, Kubernetes, GitHub Actions CI/CD, ArgoCD, Terraform, CloudWatch, Prometheus, Grafana, Kafka, Jenkins

Contact this candidate