
Agas Yanpratama
Senior Software Engineer, Ads Ranking & Search Systems
Senior backend engineer with 6 years of experience, 5+ in ads and search. Search-ads ranking at Tokopedia (TopAds), then search-ads delivery at TikTok. Now ads ranking and partner monetization at Wego, promoted to Senior within 12 months. Ships ranking and scoring models, A/B experiments, and ads services in Java, Go, and Ruby on Rails. Takes work from tech plan to dark launch to post-deploy monitoring. Directs AI coding agents (plan first, tests and CI as gates, every diff reviewed); builds LLM agents with cost caps and human approval.
- in ads and search
- 5+ yrs
- TopAds CTR per initiative
- +1–8%
- sponsored slots per mobile search
- +50%
- new service, zero to production
- 100+ RPS
Experience
Projects
Logistics operations platform and company website
Jul 2026 – Present
Internal back-office and public website for an Indonesian logistics company.
- Directed AI coding agents (Gemini CLI, OpenCode) to build the operations platform in 19 days: 73 OpenAPI routes, 18 tables, role-based access for 8 modules. Reviewed every staged diff before commit.
- Wrote the agents’ rules (AGENTS.md, plan-first workflow, architecture, security, UI) and their gates: 133 Vitest cases, dependency-boundary checks, headless-Chrome end-to-end tests, CI.
- Cut ~3,700 lines of dead code and boilerplate, plus 11 unused dependencies, with the agents; npm audit dropped from 4 findings to 0.
- Moved the company website from Django to static Astro on Cloudflare Pages; staff edit Markdown and an LLM step turns it into typed JSON.
Stack
TypeScript, Astro, Hono + Zod OpenAPI, Drizzle ORM, Cloudflare Workers, D1 & Pages, Tailwind CSS + daisyUI, Vitest, GitHub Actions
Ground Truth (Wego hackathon)
Oct 2026
Built at WegoHack 2.0, Wego’s internal hackathon. Opened 15 of the team’s 21 merged PRs, built with Claude Code.
- Designed LLM agents that check hotel listings (2 spotters, an investigator, 3 reviewers); a 50-hotel run cost USD 0.41.
- Gated every fix on a person’s approval; code verifies each number an agent quotes and caps a job at USD 1 and 270 s.
- Compared models on a 37-case golden set: DeepSeek V4.1 Flash found 36 to GLM-5.3 Flash’s 13 and ran ~8× faster.
Stack
TypeScript, Next.js, Vercel AI SDK & AI Gateway, DeepSeek V4.1 Flash, PostgreSQL (Neon), Vitest