Thoughts

Software engineers often look at AI and assume their experience is about to be replaced. It is not. Microservices and AI agent systems use the same two architecture patterns, orchestration and choreography. What changed is the actor: a service is deterministic, an agent is probabilistic, and every difference between the two worlds comes from that.

6 min read

Harnesses like OpenClaw, Hermes Agent and Deep Agents are built for a general assistant. A harness has four parts: tools, memory, control flow and evaluation. A specialised agent still needs all four, but a narrow slice of each.

2 min read

Lanjutan dari hitungan balik modal sebelumnya. Rupiah yang melemah memajukan balik modal dari bulan ke-7 ke bulan ke-5, model kecil seperti DharmaOCR mengalahkan API frontier di bidang yang sempit, dan model lokal diperbarui hampir tiap bulan.

3 min read

Dua alasan LLM on premise makin masuk akal: data sensitif tidak pernah keluar dari mesin sendiri, dan biaya tidak ikut meledak saat rupiah melemah. Lengkap dengan rumus y = wx + b, contoh hitungan di tiga volume percakapan, dan daftar biaya yang belum masuk hitungan.

5 min read

Three things kill agent projects, and none of them is the framework: cost, hallucination and the context window. All three come from what goes into the prompt. Five techniques that cut the bill on one project by about 40% by sending the right context instead of more of it.

3 min read

LangGraph solved a real coordination problem in 2024. Now that tool calling and agent-to-agent messaging have open protocols, the graph starts to feel like a flowchart wrapped around an intelligent model.

2 min read

Agent communication was standardised in the late nineties and failed commercially for being too formal. The three camps today, routers, LLM routing and protocols like A2A and MCP, are rediscovering what FIPA got right.

2 min read

Nvidia's speech-to-speech model would not run on an M4 MacBook Pro, so I rented an A40 from RunPod. Managing rental GPUs, cost per minute and knowing when to tear them down, is a skill of its own.

1 min read

Wired into Telegram on a sandboxed VM, it answered a finance-team question across business, operations and technical dimensions at once. Also why this kind of tool stays in a virtual machine.

1 min read

The protocol buffer already describes the domain, so redeclaring the same fields in Go duplicates the model. Referencing and embedding the generated types instead, and what that costs.

4 min read

The framework from a talk on AI in the workplace: start from expensive business problems rather than trends, match cloud, on-premise or hybrid to your constraints, and implement when the business impact is clear.

1 min read

Implementing a vector database on OpenSearch with TF-IDF in Python: what a vector store is for, how the index is built, and how the search behaves.

12 min read

How a marketing broadcast at Sorabel went nine times faster once the work was split across goroutines with bigger payloads, and what that taught me about batching.

1 min read