class JoharBayu:
def __init__(self):
self.role = "AI / Agentic Systems Engineer"
self.location = "Indonesia 🇮🇩"
self.building = ["Multi-step LLM agents", "RAG pipelines", "Tool-calling orchestration"]
self.backend = ["Go", "Python", "Node.js"]
self.frontend = ["Next.js", "React"]
self.motto = "Ground every answer in a source. No hallucinations."
def status(self):
return "👻 Focusing — shipping production AI agents."- 🔭 Building production LLM agents — multi-step tool calling, retrieval-augmented generation, and streaming
- 🧠 Obsessed with agent reliability: grounded citations, tool-call repair, graceful provider fallbacks
- ⚙️ I orchestrate agents hand-rolled on the Vercel AI SDK — schema-validated tools, not black-box chains
- 🌱 Going deeper on multi-agent systems, MCP, evals, and MLOps
A production regulatory & trade-compliance research agent (private)
An autonomous, multi-step tool-calling agent that answers complex domain questions by routing across several data sources and grounding every claim in a retrieved citation — no answer from prior knowledge.
| Capability | How it's built |
|---|---|
| 🛠️ Multi-step tool calling | 8 schema-validated tools + step budget, orchestrated on the Vercel AI SDK |
| 📚 RAG + grounded citations | every factual sentence must cite a retrieved source — hard zero-hallucination gate |
| 🔁 Self-healing tool calls | experimental_repairToolCall regenerates bad args against the Zod schema and retries |
| 🌊 Streaming SSE pipeline | two-phase retrieval (query-rewrite → search → select → analyze) streamed to the UI |
| ⚡ Speculative prefetch + caching | warm upstreams the moment inputs resolve; multi-layer TTL caches dedupe in-flight calls |
| 🧱 Provider-quirk resilience | recovers when the model leaks a tool call as plain text instead of a structured call |
Implementation details, prompts, and infrastructure are proprietary — described here at a high level only.
🤖 AI · LLMs · Agents
💻 Languages
🧰 Frameworks & Runtime
📊 Data Science
🗄️ Databases & Infra
⭐️ From bayujo — building agents that don't hallucinate.