Wen-Kai "SansWord" Huang
"A data-platform builder applying senior engineering discipline — measurement, cost, judgment — to AI-assisted engineering and LLM-powered systems."
Senior Software Engineer with 9+ years of experience designing and operating large-scale distributed systems and global data infrastructure on AWS and GCP, serving ~900M monthly active users across US, APAC, and Europe. Specialized in high-throughput, fault-tolerant pipelines (>20K msgs/sec), scalable data platforms and cloud-native migrations. Proven technical leader and mentor recognized for cross-functional collaboration and driving large-scale impact across globally distributed teams. Since June 2025, on a self-directed build year: designing, shipping, and publicly measuring LLM-powered systems and agentic development workflows (harness engineering) — applying the same measurement discipline (cost, model comparison, eval design) to AI work — and co-organizing AI展 (aitian), a weekly AI-application meetup that has run unbroken since Jul 2026 to 40–50 engineers a session.
- Led end-to-end delivery of Yahoo's data infrastructure evolution to GCP, building a new GCP-native ingestion platform on Cloud Composer, Airflow, and Dataproc serving US, APAC, and Europe.
- Designed the platform behind splitting Yahoo's centralized ETL pipeline into per-org, GCP-native pipelines for orgs including News, Sports, and Finance: set the boundary between the shared core schema and each org's own hydration, and built one shared module deployed from a reusable Terraform template, so each org owns its cost, freshness, and release cycle — driven by the orgs' need to own their costs separately and Yahoo's move off on-prem data centers. Rebuilt the stack natively (Oozie → Cloud Composer, Pig → Spark on Dataproc, HDFS → GCS/BigQuery) with tiered bronze/silver/gold tables that reduced dashboard data scanned from 725 GB to 121 KB and cut storage/query cost by 70%.
- As tech lead of a 6-engineer team, took it from design to production: split the architecture into four study areas, took the throughput-critical piece myself, aligned the orgs on what the central team would and wouldn't own, and agreed the hydration interface with the metadata team. Migrated Sports first, running the old and new pipelines side by side with automated acceptance checks defined with the Sports team before switching over — data freshness from 3+ hours to 30 minutes, new fields shipped in 1 week. When a silent job failure hit before a Sports standup, fixed it within the hour, shared the post-mortem in their channel, and shipped job-completion alerting tied to the freshness SLO.
- Drove an on-call reliability initiative, resolving root causes and improving monitoring to reduce open incident tickets from 1,000+ to zero; established practices subsequently adopted by the broader on-call team.
- Led cross-functional collaboration between ML research and engineering to productionize ranking models, co-inventing a patented salient entity algorithm (US11803605B2), and integrating ranking models into production.
- Drove early AWS adoption across Yahoo engineering, building blue/green deployment and auto-scaling patterns via ECS that were later adopted by global teams.
- Contributed to migrating 80 TB from house object storage to S3 with zero downtime, implementing dual-write/dual-read patterns to ensure data consistency and safe cutover.
- Scaled Pigeon, a high-availability Apache Pulsar-based message bus sustaining 20K msgs/sec with data consistency guarantees; drove migration from Apache ActiveMQ and improved MTBF from 3.5 days to 90 days (~26×).
- Core developer of Cupid, a fault-tolerant multi-tenant discount service; invented a flexible JSON-based rule engine empowering teams to independently configure promotions — driving 10K/day coupon redemption growth and boosting annual sales events.
The through-line below is measurement and judgment — applying senior engineering discipline (cost analysis, model comparison, security review, design tradeoffs) to AI-assisted work, and documenting the work openly so it stands as evidence. Three properties back this up: a public blog on the skeptic-to-shipping journey, a learning-in-public notes repo with measured cost/model studies, and the shipped projects below.
A semantic query gateway: clients send {want, where} — their own field names plus a plain-English filter — against a backend whose schema they never see. The gateway semantically resolves fields, compiles the filter to a validated predicate AST (the injection boundary — raw NL→SQL is never executed), runs it on Postgres, and answers in the client's own keys with an interpreted echo. Production-minded AI plumbing: two-part resolution cache, prompt-cache layout, confidence gate with refusal, adversarial SQLi/prompt-injection review — de-risked first by a multi-vendor resolution spike (~100% on a 52-case benchmark across 3 vendors / 9 models). Python/FastAPI gateway on Fly.io; Next.js playground with request-transparency panels + slides on Vercel.
Shipped a live full-stack product end-to-end: Next.js 16 / React 19, Postgres (Neon) + Prisma, NextAuth — with per-PR preview deploys on isolated DB branches, post-deploy smoke tests, and agentic Claude Code workflows backed by a maintained cost ledger and devlog. The product is a platform-independent identity-verification service: people prove which social accounts are the same person (anti-impersonation) and keep that proof publicly verifiable even if an account is suspended — no platform API access; verification relies on publicly auditable posted content.
A browser-only multitrack player for learning an instrument part by ear. Runs six-stem source separation entirely client-side — onnxruntime-web with htdemucs_6s on WebGPU at 8.4× realtime (23.9 s for a 200 s track) — then solos one instrument against an A–B loop driven on the audio thread so all six stems stay sample-locked. The audio never leaves the machine: no upload, no server, no build step. Output is held to the native Demucs pipeline by a parity harness (0.98–0.997 correlation per stem), backed by 81 dependency-free unit tests, zh-TW/English i18n, cookieless analytics, and per-PR preview deploys. Shipped v1.0 → v1.8 in nine days.
Real-time smart-cube solve analyzer with BLE ingestion, live 3D rendering, phase detection (CFOP/Roux), and opt-in Firestore cloud sync. Deliberately picked an unfamiliar stack to stress-test AI-assisted workflows on greenfield code.
I picked the sources — a Computerphile video, then the 1995 Cryptologia paper behind it — and Claude Code turned each one into Python attacks on an Enigma M3 cipher. Every step is measured on the same laptop: a known-plaintext crib attack in 4.4 min, and a ciphertext-only staged attack (Gillogly's 52-trial ring recovery plus a trigram-scored plugboard search) in 4.4 s, against ~4.1 h for plugboard-free brute force and ~610,000 years with a plugboard. Re-measuring every number caught an overclaim — the fast ring recovery isn't always letter-perfect — and token usage was measured from the session transcript (~58% of one Claude Pro 5-hour limit). Published as a bilingual poster (EN / 繁中) with code and a prompt-by-prompt development journal.
How I keep an AI-built project on track: a lazy-loaded document tree — a small auto-loaded index pointing to topic docs, a two-tier split between maintained and historical docs, one-fact-one-place, and using a fresh-context agent to catch where the docs have silently drifted from the code. Grounded in the real CLAUDE.md / docs/ trees of guasi and sans_cube, and productized as sans_doc_tree — a reusable GitHub template of the structure.
My writing home: building software, and building it with AI. Highlights — My Tiny AI Bootcamp (the skeptic-to-shipping story), a worked Leaderboard System Design study guide, and agentic-coding pieces like The Document Tree (above).
AI-round mock interview kit. Multi-mode Claude Code agent — a root CLAUDE.md mode-routes between interviewer and developer modes on a single repo. Measured per-mock API cost (~$19–$26) across Pro / Max / raw-API tiers; includes a feedback rubric, a full real-run example, and a security-aware CLAUDE.md audit recipe for safe repo onboarding.
Co-organizer and moderator of a weekly online AI-application meetup — members ("aitians") demo what they're building and trade AI workflows. Started from zero in Jul 2026 and has run every Tuesday since, unbroken — 40–50 attendees per session (peak 107), bilingual EN/ZH. I recruit and moderate the speaker roster — engineers from AWS and Netskope, a Vector Institute AI researcher, a UIUC informatics PhD, and AI-startup founders — set each week's theme, and run the open-discussion segment. Also built the meetup's bilingual portal end-to-end at aitian.dev: Markdown/YAML content validated and built to JSON by CI, with a PR-based contribution flow so moderators and speakers own their entries — shipped with agentic Claude Code workflows. Presented sans_schema at the opening session.
Co-invented methods to identify salient entities in articles at scale and apply them in production recommendation systems at Yahoo.
| Languages | Python, Java, SQL, Shell |
| Cloud — GCP | Dataproc, Cloud Composer, Dataflow, Airflow, BigQuery, GCS |
| Cloud — AWS | S3, ECS, ELB, ElastiCache |
| Infrastructure | Terraform, Kubernetes, Docker, Redis |
| Data Processing | Spark, Hadoop (HDFS), Apache Pulsar, Kafka |
| Data Modeling | Medallion architecture (bronze/silver/gold), shared core schema with per-org extensions |
| AI / LLM Engineering | Multi-vendor LLM integration (LiteLLM), prompt caching & cost engineering, eval design, guardrails (validated-AST query compilation, prompt-injection hardening), agentic workflows (Claude Code), harness engineering |