Plank AI Accelerator
Become a world-class FDE
AI is eating software, fast, and the future of programming is changing dramatically. We are training a new generation of forward-deployed AI engineers (FDEs) who turbo-charge their skills for this new era, rather than be left behind.
The program
A competitive one-year program
Designed to help ambitious early-career engineers develop deep expertise in building end-to-end GenAI apps. Spots are limited to keep the bar high.
One year, fully supported
A competitive program with limited spots, and a monthly financial award for the full duration.
Real silicon-valley projects
Build end-to-end AI apps on Plank project teams, with a front-row seat to how fast-growing companies ship AI.
A guaranteed role
Complete the program and join Plank as a core-team AI engineer.
How it runs
Three phases
Phase 1
Train
Intensive training by technical leaders on the building blocks of GenAI apps: swarm agent pipelines, voice agents, LoRA adapters, and on-device models.
Phase 2
Build
A week-long hackathon where you put it all together and ship two end-to-end GenAI applications.
Phase 3
Forward-deploy
Embed as a forward-deployed engineer on real-world AI projects with Plank’s commercial partners, scoping and shipping to production alongside senior executives and engineering leaders.
The curriculum
What you will master
Six modules that take strong full-stack engineers from the agentic development workflow all the way to forward deployment.
Module 1
The agentic development workflow
AI-paired development across Claude Code and Codex, agent skills and sub-agents, the harness-driven AI SDLC, multi-project workflows on git worktrees, and custom MCP servers.
Module 2
Building AI applications
The anatomy of a production AI app: the Messages API and sampling controls, tool use and call chaining, full-stack streaming chat, generative UI, and context and memory management.
Module 3
Agents, RAG & retrieval
Agents from scratch and with frameworks, multi-agent orchestration, modern agentic retrieval over real corpora, and browser, desktop, and scraping automation.
Module 4
Multimodal: voice & video
Real-time voice agents with streaming speech-to-text and text-to-speech, interruption and turn-taking, plus video understanding and the video-generation landscape.
Module 5
Production, evals & deployment
Structured extraction, durable workflow orchestration, the testing pyramid and evals, tracing and prompt management, cost optimization, and deployment with alerting.
Module 6
Forward deployment
What sets a Plank FDE apart: embedding with a customer team, scoping architectures with stakeholders, designing for reliability, integrating real systems of record, and making pilots stick in production.
The full curriculum
Everything you cover
An index of the ground every Plank engineer covers, each topic paired with a hands-on build and evals. This is the depth behind a Plank pod.
The agentic development workflow
- AI pair-coding across Claude Code and Codex
- CLAUDE.md project and folder conventions
- Claude Code skills and custom slash commands
- Sub-agents for parallel review and complex tasks
- Event hooks: pre-commit, post-edit, and built-in commands
- The AI SDLC: ticket, design, code, test, PR, deploy, monitor
- Adversarial code review with the Codex plugin
- Where humans add value versus where AI does
- Linear, GitHub, and Figma MCP for tickets, PRs, and design specs
- Multi-project workflows on git worktrees with branch conventions
- Parallel agents: shipping several features in one session
- MCP architecture: tools, resources, and prompts; stdio versus SSE
- Building custom MCP servers in TypeScript
- Spec-driven, review-driven development
Building AI applications
- Provider SDKs and the Messages API
- Conversation structure: roles and content blocks
- Model selection: reasoning versus balanced versus fast and cheap
- Sampling parameters: temperature, top-p, max tokens, stop sequences
- Streaming versus non-streaming with server-sent events
- Prompt engineering: system prompts, few-shot, chain-of-thought
- Tool use: definitions, JSON-Schema inputs, the call loop
- Multi-turn tool loops and human-in-the-loop confirmation
- Chaining LLM calls into multi-step workflows
- Full-stack AI chat: route handlers and server-side streaming
- The Vercel AI SDK and rendering streamed markdown
- Generative UI: LLM-driven JSON to React components
- Constrained component catalogs with Zod and progressive rendering
- Token budgets, truncation, sliding windows, and compaction
- Long-term memory: fact extraction, retrieval, and decay
- Memory storage across Redis, PostgreSQL, and vector stores
Agents, RAG & retrieval
- Agent versus tool use; the observe, think, act loop
- ReAct, plan-then-execute, and Reflexion patterns
- Building an agent from scratch with guards and stop conditions
- Multi-agent patterns: orchestrator, pipeline, swarm, and debate
- Agent handoffs and shared context
- The Claude Agent SDK, LangGraph, and CrewAI
- Why brute-force RAG fails, and when it is enough
- Agentic retrieval: search, read, refine, cross-reference
- The virtual filesystem pattern: ls, cat, grep, find over a corpus
- Embeddings and vector search with Voyage AI and pgvector
- AI scraping: schema-based extraction from messy HTML
- Browser automation with Playwright and vision, Firecrawl, Browserbase
- Desktop and computer use
Multimodal: voice & video
- Speech-to-text with Whisper, Deepgram, and AssemblyAI
- Text-to-speech and voice cloning with ElevenLabs and Cartesia
- Real-time voice: streaming, voice activity detection, turn-taking
- Interruption handling and latency budgets for natural conversation
- Audio capture with MediaRecorder and the Web Audio API
- Video understanding: frame extraction and vision analysis
- The video-generation landscape and APIs
Production, evals & deployment
- Structured output: tool-use as a typed schema
- Zod validation with retry-and-feedback loops
- Document and data extraction pipelines with confidence scoring
- Durable workflow orchestration: checkpoint and resume
- Fan-out, fan-in, and human-in-the-loop approvals
- The testing pyramid: unit, integration, end-to-end, and evals
- Mocking LLMs, snapshot and regression testing
- Evaluation: LLM-as-judge, golden datasets, statistical scoring
- Tracing and observability with Langfuse
- Prompt management, versioning, and A/B testing
- Cost optimization: prompt caching, model routing, token budgets
- Deployment, alerting, rate limiting, and PII handling
Forward deployment
- Embedding with a customer team
- Scoping architectures with stakeholders
- Integrating real systems of record
- Designing for reliability: fallbacks and circuit breakers
- Product thinking: mapping business goals to AI features
- Build versus buy, and managing expectations
- Making pilots stick in production
- Documenting repeatable patterns your team owns
What you build
Production-grade projects
Every topic is paired with a hands-on build. You ship more than twenty end-to-end projects, each evaluated, before you ever embed with a customer team.
Streaming chat app
A full-stack AI chat app with tool calls, generative UI, and a memory system, deployed to production.
Generative-UI dashboard
An interface that renders live React components from natural language, on a safe, constrained component catalog.
Agentic retrieval engine
Search, read, and cross-reference a document corpus, benchmarked head to head against brute-force RAG.
Multi-agent research analyst
Orchestrated agents that combine internal retrieval and live browsing into a single sourced report.
Real-time voice assistant
Streaming speech-to-text, reasoning, and text-to-speech, with interruption handling and turn-taking.
Structured extraction pipeline
Typed, validated extraction from invoices, resumes, and emails, with retries and confidence scoring.
Durable document workflow
An orchestrated pipeline with per-step retries, parallel fan-out, and human-in-the-loop approval.
Production hardening
An app taken live with an eval suite, tracing, prompt versioning, and cost alerting.
Architecture capstone
A complete system design and pitch against a real startup brief, with tradeoffs made explicit.
Tools & stack
The stack you ship on
You train hands-on across the modern AI stack, vendor-neutral, so you are productive in a real codebase from day one.
Agentic dev
Models & SDKs
App & UI
Agents & RAG
Retrieval & browser
Voice & video
Orchestration & ops
Cohorts
Apply to a cohort
FAQ
Common questions
What is a forward-deployed engineer (FDE)?+
A forward-deployed engineer embeds directly with a customer team to take AI from pilot to production: scoping the architecture, integrating with real systems of record, and shipping alongside in-house engineers. It is the role behind FDE as a service, and the role the Plank AI Accelerator trains you for.
What is the Plank AI Accelerator?+
The Plank AI Accelerator is a competitive one-year program that trains ambitious early-career engineers into world-class forward-deployed engineers. It runs in three phases: intensive training on the building blocks of GenAI apps, a build phase shipping end-to-end applications, and a forward-deploy phase on real projects with Plank's commercial partners.
Is there a job after the program?+
Yes. Complete the program and you join Plank as a core-team AI engineer, embedding with the companies Plank serves through FDE as a service. The program is fully supported, with a monthly financial award for the full duration.
What does the curriculum cover?+
Six modules: the agentic development workflow, building AI applications, agents and retrieval, multimodal voice and video, production and evals, and forward deployment. Every topic is paired with a hands-on build, and engineers ship more than twenty end-to-end projects across the modern AI stack, vendor-neutral.
Who should apply?+
Ambitious early-career software engineers who want to become forward-deployed AI engineers rather than be left behind as AI rewrites how software is built. Spots are limited to keep the bar high.
Turn your skills into a career in AI
Join the Plank AI Accelerator and graduate as a world-class AI engineer, with a role at Plank waiting for you.