Context retrieval agent
Designed and built the orchestrator agent that, given a user query, plans and executes the right sequence of search / explore / file calls to assemble the context downstream AI features consume.
Open to AI-first teams
📍 Fremont, CA
Arky Yang
Core engineer building an enterprise-scale context platform for AI agents — end to end, from zero to release.
In a large enterprise, context is scattered across 10,000+ repos. AI made writing code in any one repo fast — so the bottleneck moved to cross-service context and coordination. We build the engine that closes that gap — a small team of ~10 engineers that moves faster than any other: deep cross-repo code search, where an organization's business logic actually lives, fused with docs and other sources into the comprehensive context an agent needs to act.
Now
🏗 Context platform for AI agents · Atlassian Rovo Dev
🛠 Rederive · a "Roblox for education" platform (building)
🐻 Cedartale · voice-cloned AI storytelling (shipped)
01 — Summary
The engineer who holds both sides of the agent boundary
I built an enterprise context engine for AI agents and took it from zero to external release — the thing an external-facing AI chat product and external customer launches now run on, and the SDLC agent products our company is building will rely on it, too. I worked as the connective engineer across 10+ partner teams to turn diverse business needs into one coherent, agent-facing platform. On the side I build Rederive and Cedartale, which keeps me sharp on both the agent-side ergonomics and the production-system constraints — at the same time.
End‑to‑end
Owned the context engine — itself an agent, backed by powerful cross-repo code search — across API, auth, ops & UI
10+ teams
Partner teams coordinated for one external onboarding journey
Zero → External
From zero to external release, powering live customer launches
Agent-first
MCP-exposed platform + CLI agent full-context mode
02 — Selected Work
Software Engineer, Code Intelligence
Cross-repo code search at scale
Designed and implemented the public API contract — versioning, error semantics, pagination, permission semantics, MCP-friendliness — that every consuming agent, internal service, and external customer integrates against. Exposed through MCP for internal agents and a CLI agent's full-context mode.
External tenant & host onboarding
Took code search from one source host to multi-host (one third-party host launched, another in progress). Coordinated 10+ partner teams — source control, provisioning, billing, edge gateway, AI product — into a working end-to-end onboarding journey for external customers, including GitHub repo index onboarding.
Permission & auth across hosts
Multi-tenant permission resolution, account-ID-to-tenant mapping for third-party code, and a governed service-to-service onboarding contract so downstream callers integrate through a documented process rather than ad hoc.
Production & operational readiness
Multi-tier rate-limit policy (per-client × per-user × per-operation) distinguishing humans, internal bots, external clients, eval traffic, and load tests, plus a generous edge-gateway ceiling. Experience-id propagation for traffic attribution, an integrator troubleshooting runbook, and trace-first debugging for "why did the agent get nothing?" failures.
Code search & chat frontend
Built and shipped the user-facing surfaces for code search and the AI chat experiences that consume it (React/TypeScript) — so the same context engine powers both programmatic agents and end-user product flows.
Raised the team's AI leverage
Improved project-level agent harnesses and shared context (skills, prompts, runbooks, MCP setups) so teammates can pick them up directly, and share AI-assisted workflows and tooling discoveries (Claude Code, MCPs, custom skills) with the team on a regular cadence.
03 — Side Projects
Both sides of the agent boundary
🛠
In progress · unreleased
Rederive
A "Roblox for education" platform with an agent harness for creation.
Anyone with an AI agent can author courses in their domain, and a Socratic teaching agent (also built by me) guides learners through them. The authoring side is a BYO-agent, YOLO-mode pipeline that goes from one prompt to a fully animated, narrated video or interactive course.
- Autonomous long-horizon execution
- BYO-agent contracts (MCP-style, at the agent level)
- Zero-to-one creative pipelines
- The pedagogy & UX of a teaching agent
Same problem shape as agent-coding harnesses — applied to education.
🐻
Shipped
Cedartale
Voice-cloned AI storytelling app for kids.
A parent's own voice can read the story. Founder, sole engineer, sole product designer — the whole loop from idea to a shipped consumer app.
- Voice cloning & generative storytelling
- Consumer product, end to end
- Where I think & write publicly about raising kids alongside generative AI
Founder · sole engineer · sole product designer.
Download on the App Store ↗
04 — Stack
What I use day to day
Backend
Kotlin / SpringPythonGo
RedisGraphQLOpenSearch
Infra & platform
Edge / API gatewayService-to-service auth
User authEvent-streaming
Container deploy pipelines
AI
Claude CodeRovodev CLI
CursorWindsurf
MCPs & skills
Observability
Distributed tracingProduction log platforms
Time-series dashboards
Frontend
ReactTypeScript
Admin & portal UIs
05 — Education & before
Chemistry, then computer science
MS, Computer ScienceUT Austin
BS, Chemistry (Highest Honors)UC Berkeley
Chemistry, Resnick FellowCaltech
ChemistryHKUST
Visiting Student, Organic ChemistryUC Santa Barbara
CS coursework: Android, deep learning, computer vision, optimization, front-end. Pre-CS research: 2D organic–inorganic hybrid perovskite materials for optoelectronics (Harry Atwater Group, Caltech), copper nanowire synthesis for transparent electrodes (Peidong Yang Group, UCB), polymer solar cells (He Yan Group, HKUST). Publications include Giant Enhancement of Photoluminescence Emission in WS2–Two-Dimensional Perovskite Heterostructures.
Earlier — SWE Intern (2021)
- React/TypeScript + Jest reusable card UI components for distributed services.
- Event-driven Spring Boot/Kotlin/GraphQL backend for default-component provisioning on site creation.
- In-product help search experiments (Algolia, React Context & Hooks) with content design.
Outside engineering (it shows up in the work)
- President, Caltech Entrepreneurship Club (2017–2020) — alumni-CEO panels, the FLOW/Rockey Fund program for young tech companies, Idealab field trips.
- Plug and Play (2019) — analysed pitch presentations across verticals.
- Calibr at Scripps Research (2018) — research & market landscape of new depression meds.
- Mom of a two-year-old — an excellent forcing function for thinking carefully about AI.