方寸 Portal

AI Engineer Portal

Your personal operating system for career transition.

Private mode

Learn

Structured progression, not fragmented browsing.

Paths for sequential study. Reference articles for deep dives on specific topics.

Learning Paths

Python for AI Engineers

Build practical Python fluency for APIs, pipelines, evaluation scripts, and debugging in applied AI systems.

0% complete
beginner6h
Python runtime habits that matter in AI work
Async patterns for AI workloads
Data pipelines for AI ingestion
Open path

LLM App Foundations

Learn the product and systems primitives behind useful LLM applications.

0% complete
intermediate16h
Prompt, context, tools, and memory
Request lifecycle of an LLM feature
Guardrails and structured outputs
Open path

RAG Systems

Build retrieval systems that are explainable, measurable, and debuggable.

0% complete
intermediate18h
Document ingestion and chunking
Embedding and vector storage
Retrieval patterns
Open path

AI Agents and Tools

Understand when agentic patterns help and how to keep them safe.

0% complete
intermediate14h
Tool design for LLM agents
Planning and reasoning loops
Multi-agent orchestration
Open path

Evaluation and Observability

Measure behavior before scaling usage and complexity.

0% complete
intermediate12h
Evaluation metrics for LLM outputs
Building eval harnesses
Tracing and logging AI requests
Open path

AI Deployment and MLOps

Make AI systems operable beyond the first successful demo.

0% complete
advanced18h
Containerizing AI applications
Scaling LLM inference
CI/CD for AI features
Open path

AI Engineer Interview Readiness

Translate execution experience into role-aligned interview performance.

0% complete
intermediate10h
Tell the transition story clearly
System design for applied AI
Python and backend refresh under pressure
Open path

AI Safety and Guardrails

Build defense-in-depth for production AI systems. Covers prompt injection defense, content filtering, PII protection, red teaming, and governance frameworks.

0% complete
intermediate12h
The AI safety landscape for application engineers
Prompt injection defense
Content filtering and output validation
Open path

Context Engineering

The meta-skill of 2026: designing the entire information environment an LLM operates in. Goes beyond prompting to manage memory, retrieval, tool results, and context window budgets.

0% complete
intermediate14h
From prompt engineering to context engineering
Context window budgeting
Memory architecture for AI applications
Open path

Reference Articles

architecture

A clean handoff between ingestion and retrieval

The indexing layer should emit chunks and metadata in a form the retrieval layer can trust without guesswork.

concept

A decision guide for prompting, RAG, and fine-tuning

Pick the right technique based on knowledge freshness, control needs, and operational cost.

concept

A portfolio roadmap for the AI engineer transition

Pair learning paths with portfolio artifacts so each month produces visible proof, not just study notes.

architecture

A practical schema for AI request traces

Store enough detail to debug context, prompts, outputs, latency, and cost without overwhelming yourself.

deployment

A/B testing and safe rollout for model deployments

How to safely deploy new model versions using traffic splitting, quality metrics, and staged rollout — without full-deployment risk.

agents

Agent architecture patterns compared

A practical comparison of ReAct, Plan-Execute, and Tree-of-Thought agent architectures with a decision matrix covering latency, token cost, reliability, and complexity trade-offs.

python

Async patterns for high-throughput AI data pipelines

Producer-consumer queues, backpressure, fan-out with semaphores, and the operational tradeoffs between asyncio and multiprocessing for AI workloads.

architecture

Choosing between sync APIs and background jobs

Move heavy ingestion and evaluation work off the request path once latency or reliability starts to matter.

architecture

Chunking strategies for product-grade retrieval

Choose chunk sizes and boundaries to preserve meaning, support citations, and improve ranking.

architecture

Citations are a UX feature, not a footnote

Citations help the user calibrate trust, inspect weak answers, and continue research rather than starting over.

evaluation

Cost-aware evaluation strategies

How to build evaluation systems that stay within budget by combining cheap deterministic checks, sampling strategies, and expensive judge models at the right points in the pipeline.

architecture

Debug retrieval before changing prompts

Weak answers often come from bad context selection, so inspect retrieval traces before rewriting prompts.

deployment

Deployment patterns for LLM applications: from demo to production

The engineering decisions that separate a working demo from a production AI feature: reliability, cost controls, rollout strategy, and degradation planning.

architecture

Deployment readiness for LLM features

Production readiness means secrets, retries, caching, tracing, and rollback plans are all intentional.

architecture

Design your portal like a product, not a notebook

Treat content structure, navigation, and progress loops as product decisions that should reduce friction over time.

architecture

Designing a model fallback chain: resilience without complexity

How to build provider fallback chains that improve reliability, reduce cost on retry, and degrade gracefully — without making your codebase a tangle of exception handling.

python

Error handling patterns for AI service integrations

How to classify provider errors, implement retry policy correctly, build fallback chains, and surface meaningful error context without leaking internals.

concept

Evaluation metrics that actually help iteration

Use metrics that point to specific failure modes rather than one vague quality score.

architecture

Failure modes worth logging explicitly

Log missing context, schema mismatches, provider failures, judge disagreement, and human override reasons.

evaluation

Golden dataset curation and maintenance

How to build, curate, and keep a golden evaluation dataset useful over time — including seeding from production traffic, avoiding common data quality traps, and retiring stale cases.

architecture

How to choose between OpenAI and Anthropic for a production feature

A practical framework for selecting a provider based on capability, reliability, cost, and integration fit — not hype.

architecture

How to review an AI feature after launch

Post-launch review should inspect traces, evaluation drift, user pain points, and operational cost together.

concept

How to think about provider lock-in

Provider portability matters most at the boundary layer where schemas, retries, and cost controls are centralized.

concept

Human review is part of the system

Human checkpoints are a strength when confidence is low or business impact is high.

architecture

Hybrid retrieval: when to use it and how to tune it

Combine vector and keyword search using RRF, understand when each method wins, and tune the balance for your query distribution.

concept

Interview prep should mirror shipped work

The best prep material points back to projects, metrics, and tradeoffs you actually worked through.

evaluation

LLM-as-judge: design principles and failure modes

How to design reliable LLM judge prompts, calibrate them against human labels, and avoid the biases that make automated evaluation misleading.

architecture

Latency budgets shape product design

Perceived speed depends on retrieval, provider calls, streaming behavior, and how the UI acknowledges work in progress.

python

Managing LLM response variability in production

Practical patterns for parsing, validating, and recovering from the structured output failures that LLMs produce in production.

agents

Measuring agent effectiveness

Practical evaluation approaches for production agents covering task completion rates, cost efficiency, reliability metrics, A/B testing, and offline eval harnesses.

architecture

Observability for AI requests

Trace context assembly, provider calls, outputs, and failures so AI bugs become debuggable engineering work.

deployment

Observability patterns for production LLM services

What to measure, trace, and alert on in production AI services — beyond basic uptime and latency.

evaluation

Observability stack for production LLM applications

What your traces, logs, and dashboards need to contain to debug AI quality issues in production — and how to structure the data pipeline from call to alert.

deployment

OpenTelemetry for LLM applications: what to trace and why

How to add span-level tracing to LLM pipelines using OpenTelemetry, and which attributes make traces actionable for debugging slow or low-quality responses.

concept

Personal knowledge bases need freshness rules

Every content library should say what is stable, what expires, and what gets reviewed on a schedule.

architecture

Production RAG: the operational checklist before you launch

The eight operational concerns that separate a demo RAG system from one you can run reliably: caching, fallbacks, monitoring, cost, freshness, access control, evaluation, and latency.

concept

Prompt evaluation needs qualitative notes too

Scores alone miss tone, structure, and user trust; keep notes that explain why a version won or failed.

concept

Prompt templates should behave like contracts

Treat prompts as explicit interfaces with versioning, variable boundaries, and review criteria rather than magic strings.

concept

Provider wrappers should be boring

A good provider wrapper normalizes responses, centralizes retries, and makes the rest of your system simpler.

python

Python patterns every AI engineer should know

The specific Python patterns that matter for AI engineering: Pydantic boundaries, async concurrency, generator pipelines, and structured logging.

architecture

RAG context window management: fitting retrieved chunks into prompts

Manage the LLM context window as a scarce resource: token budgets, deduplication, relevance-ordered assembly, and the lost-in-the-middle effect.

deployment

Rate limiting strategies for multi-tenant LLM services

How to implement token-aware, multi-tier rate limiting for AI APIs — and why request-count limits alone fail for LLM workloads.

evaluation

Regression testing discipline for LLM-powered features

How to build a regression testing practice that catches quality degradation before deployment, covering dataset curation, CI integration, and triage workflow.

architecture

Retrieval metadata is a product decision

Metadata design decides what the system can later filter, cite, and explain back to the user.

deployment

Semantic caching: when to use it and how to tune the threshold

Semantic caching reduces LLM costs and latency for FAQ-style workloads. This guide covers when it helps, when it hurts, and how to calibrate the similarity threshold.

python

The async patterns that actually matter for AI workloads

Which async Python patterns produce real throughput gains for LLM-heavy code, and which common patterns introduce new failure modes.

architecture

Token budgeting: the engineer's guide to context window management

How to think about the context window as a finite resource, allocate it intentionally across prompt components, and avoid the silent failures that come from ignoring token counts.

python

Token counting and context budget management in Python

How to count tokens accurately using tiktoken, build token budget managers, and prevent context window overflow before it reaches the API call.

agents

Tool design best practices for production agents

How to design reliable, testable agent tools using JSON Schema, structured error contracts, idempotency, isolation testing, and versioning strategies.

concept

Use benchmark regressions to drive weekly work

Weekly iteration improves when regressions produce concrete follow-up tasks instead of generic worry.

concept

Use learning systems to support real projects

The portal should feed execution: learn, apply, review, and convert the result into portfolio evidence.

concept

What a strong AI project write-up includes

A strong write-up explains the problem, architecture, tradeoffs, evaluation method, and what you would improve next.

architecture

What makes a RAG system trustworthy

A practical framework for grounding, citations, retrieval transparency, and evaluation in RAG products.

concept

When to persist generated artifacts

Persist prompts, contexts, scores, and outputs when they help review, replay, or explain product behavior.

concept

Why evaluation sets should start small

A curated, trusted benchmark set beats a larger but noisy dataset when you are still learning what failure looks like.

concept

Why many agent demos fail in production

Open-ended loops hide state, cost, and failure reasons unless you add explicit boundaries and observability.