Engineering writing from people who shipped the thing.
No vendor-blog SEO. No “the future of AI is…” hot-takes. Architecture decisions, eval surprises, on-call lessons — written by the engineers who lived them.
Why LLM-as-judge failed our brokerage compliance gate.
96% pass rate sounds good. The 4% is what gets a brokerage fined. How we swapped to a deterministic rule layer + sentence classifier and got to 99.97%.
STAFF · TPS
Twelve specialist agents beat one big prompt — every time.
The architecture choice that made TPS Mobile AI auditable. Hard-typed I/O, separate eval suites per agent, and how we kept latency under 1.4s p95.
SR · GO PLATFORM
The 100-line eval runner we should have shipped in week one.
We spent three weeks on a perfect eval harness before we had three agents to evaluate. Don’t do that. Here’s the minimum viable harness, and how to grow it.
SR · INFERENCE
Streaming agents over flaky 4G — why SSE beat WebSockets.
Our first streaming protocol assumed clean WebSockets. Vietnamese retail traffic does not. The day-one design we wish we’d had: SSE with reconnect-from-offset.
SR · REACT NATIVE
Vietnamese vernacular tokenization is a research project, not a stack swap.
What we learned self-hosting Qwen-VN. Diacritics, code-switching with English, and why off-the-shelf tokenizers fail on Vietnamese financial text.
SR · INFERENCE
Our CLAUDE.md standard — how every PR ships agentic.
The repo file every CoderPush project carries. What goes in it, what doesn’t, and why we made it required for every senior engineer.
HEAD OF ENG
VPC-native Bedrock for regulated finance — a procurement guide.
What CISOs ask, what auditors want, and what to put in the data-flow diagram. Distilled from nine banking engagements.
SR · DATA
What “AI-native” actually means — on a real engineering team.
Not a hiring slogan. The five practices — codified evals, agent harnesses, prompt PRs, replay tests, ops dashboards — that separate a 2025 team from a 2019 one.
HEAD OF ENG