Skip to content

Mum, I wanna be an AI Engineer

Select theme

Factory map

M0 · merged

Inference economics

Tool loop, streaming, prefix caching, cost ledger, batch builder

stages: 7 stages
live record: synthetic transports

Tokens, prefill, decode
KV cache and batching
Prompt caching as prefix engineering
Cost levers: batch, effort, budgets, model choice
Structured outputs and strict tools
Context management: compaction, editing, memory
Interview angle: inference economics

M1 · merged

Enforced harness

Policy engine, PreToolUse hooks, durable sessions, caps, reviewer subagent

stages: 8 stages
live record: not recorded

A harness is a loop plus a policy
Hooks, permissions and modes
Subagents
Durable sessions and idempotency
Context engineering for code tasks
Failure taxonomy and caps
Interview angle: enforced harness

M2 · merged

Evals as engineering

Trace loader, labels and kappa, schema-bound judge, trust and regression gates

stages: 8 stages
live record: judge 20/20 outcome

Error analysis first
Judges and agreement
Datasets and contamination
Agent evals
CI gating and tracking
Evals for code agents
Interview angle: evals as engineering

M3 · merged

Observability and cost

GenAI spans, cost from spans, budget guard, Langfuse export, router

stages: 8 stages
live record: 9 spans · Langfuse 4.63

Traces and spans for agents
GenAI semantic conventions
Langfuse, self-hosted
Cost per task
Routing and fallback
Latency engineering
Interview angle: observability and cost

M4 · merged

Retrieval as a component

Chunker, pgvector store, hybrid RRF, reranker, metrics, cited answers

stages: 9 stages
live record: recall@10 0.78 · MRR 0.65

When not to build RAG
Chunking and embeddings
Hybrid search
Reranking and MMR
Retrieval metrics and golden sets
Citations, faithfulness and agentic retrieval
Interview angle: retrieval as a component

M5 · merged

Security and compliance

Injection detector, provenance and taint, secrets, remote MCP auth, audit log

stages: 9 stages
live record: injection denied · checklist 10/10

A threat model for agent factories
Indirect injection and provenance
Permissions, sandboxing and secrets
Remote MCP and OAuth
Red teaming with promptfoo
The EU AI Act for engineers
Interview angle: security and compliance for agents

M6 · merged

Serving and platform

Load generator, local models, K8s manifests, autoscaling, gateway

stages: 8 stages
live record: gateway 6/6 routed · $0.057

Open-weight landscape and licences
Serving internals
Running locally
Kubernetes for inference
Measuring TTFT, TPS and cost
Fine-tuning, when and how
Interview angle: serving and platform

M7 · merged

Capstone: pre-review

Rule loader, read-only reviewer, verify gate, blocker ranking, metrics

stages: 8 stages
live record: 12 PRs · recall 0.80

Review as a measurable pipeline
Conventions of a Go, TypeScript, proto and migrations repository
Reading a PR with a read-only agent
Blocker ranking from repository signals
From pre-review to what blocks prod
Interview angle: pre-review and what blocks prod

local API not running: start it with uv run miubais serve