AI Engineering Roadmap: Build 12 Core Components from Scratch
EAccelerate_42 · x · 2026-08-31
Addressing the question of what to build in AI Engineering, the author provides a practical list of projects to implement from scratch, focusing on deep understanding over reliance on abstracted libraries.
Key components include:
- Agent Orchestrator: A deterministic state machine implemented without LangChain.
- Context Assembler: Token-budgeted memory combined with retrieval and tools.
- MCP Server & Client: Raw JSON-RPC implementation, bypassing SDKs.
- Retrieval Stack: Custom implementation of chunking, BM25, dense retrieval, and reranking.
- Eval Harness: Trajectory grading and CI regression gates.
- Model Router: Intelligent routing based on cost, latency, and quality with fallbacks.
- Semantic Cache: Embedding similarity caching with hit-rate tracking.
- Guardrails Middleware: Middleware for injection detection and PII redaction.
- Streaming Proxy: SSE support with metrics for Time to First Token (TTFT) and Inter-token Latency (ITL).
- Durable Workflow Engine: Checkpoint and resume mechanisms (mini-Temporal).
- LLM Tracer: OpenTelemetry-style spans for tracing every hop.
More from coding & agent
- New MCP connector allows natural language control of Google Cloud Datastream — modelcontextprotocol · 2026-08-31
- Rolli MCP: Social media search and analytics across major platforms — modelcontextprotocol · 2026-08-31
- Copilot CLI debugs AirPods plugin dependency issues — DanWahlin · 2026-08-31
- Slide Deck Released: Introduction to Harness Design — 4310sy · 2026-08-31
- drskill released: A 'brew doctor' for checking Agent context and MCP configs — dbreunig · 2026-08-31
- Musk showcases Grok Bot workflow for building research agents — elonmusk · 2026-08-31