13 Lessons from Building Enterprise AI Agents: Self-built ETL Cuts Costs 20x
redouanea · reddit · 2026-08-04
A developer built voice and chat AI agents for technical field services and summarized 13 core practical lessons.
Tech Stack
- Orchestration & API: Python + FastAPI, LangGraph for graph orchestration, LangChain for component nodes.
- Retrieval & Observability: OpenAI embeddings and Milvus for hybrid search, ColBERTer-2 for reranking; monitored with Langfuse.
- Infrastructure: Hetzner servers combined with AWS, supporting automatic failover across multiple model providers.
Key Lesson: Own Your Data Pipeline
- Build ETL In-house Early: Initially used the commercial platform Unstructured, but it failed frequently and was costly when processing 8,000 pages of complex engineering docs. Switching to the open-source library Docling improved extraction quality and gave full architectural control.
- Massive Cost Reduction: Moving off per-page commercial pricing and running on their own GPU infrastructure cut processing costs by 15 to 20x.
- Extraction Quality Caps Performance: Bad data extraction cannot be fixed by a better retriever or model; underlying document processing quality determines the system's maximum performance.
More from coding & agent
- Seeking Free MCP Servers for Mobile App UI Design Inspiration — photographywithdc · 2026-08-04
- Musubi Tuner Fork Enables MiniMax H3 LoRA Training on 24GB VRAM — diogodiogogod · 2026-08-04
- AI Commits v3 Released: 5x Faster PRs and Local Model Support — nutlope · 2026-08-04
- Fletch: A macOS App Giving Claude Code Agents Sandboxed Workspaces — a1chapone · 2026-08-04
- Securing AI Agents: An MCP Payment Flow That Cannot Hijack Private Keys — pvdyck · 2026-08-04
- Zilla 2.0 Natively Integrates MCP: Bridging REST and Kafka Without Wrappers — jkriket · 2026-08-04