Plank AI Accelerator - 2-Month Program
The AI FDE Course
Free forward deployed engineer (FDE) training from Plank, the FDE as a service company. In 40 days, a full-stack engineer builds and ships one production AI product, WorkBot, working with coding agents throughout.
By Anusheel Bhushan, co-founder of Plank
- Duration
- 40 days of work, in 8 weeks or faster
- Format
- Individual. You build one product, plan each day's tasks, and review every pull request before it merges
- Audience
- Full-stack engineers who work in TypeScript and Node.js, at any stage of their career
- Stack
- TypeScript throughout, with no lock-in to one AI provider. Anthropic and OpenAI models and SDKs side by side. Claude Code, Codex CLI and OpenCode as coding agents. LangGraph.js, the Claude Agent SDK and the OpenAI Agents SDK for agents. The Model Context Protocol (MCP) for tools. Vendor-neutral tools for speech, retrieval, workflows and observability
- Goal
- Full-stack engineers become AI engineers
Use AI and Build AI
Use AI
How you work. Coding agents (Claude Code, Codex CLI, OpenCode) plan, write, test and review the code. You specify, steer and decide. You start with one agent and a pull request loop. By Week 5 you run a fleet of agents through a backlog, with evals and a merge queue deciding what merges
Build AI
What you ship. Models inside the product: chat and voice, agents with tools, search that answers with citations, documents read and understood, and integrations that act in other systems. Tests, evals, tracing and deployment make it hold up in production
Use AI is how you work every day. Build AI is what most days add to WorkBot.
What you learn
- Working with coding agents
- Claude Code, Codex CLI and OpenCode as your development team. Agents write the specs, pull requests and reviews. Several agents work in parallel on git worktrees, and then 50 at once on a backlog
- Chat and voice assistants
- Streaming chat on the web and in Slack, WhatsApp, SMS and iMessage. Voice that listens, speaks and handles interruptions. Screens the model composes from a catalog of components
- Agents that use tools
- Agents that search, read, browse and act, remember what matters, and hand work to each other. The same agent built on 3 frameworks, so you can choose one
- Search over company knowledge
- Retrieval that explores a knowledge base, policies and manuals, and answers with citations. You measure it against plain retrieval
- Documents, images and spreadsheets
- Structured data pulled out of PDFs, scans, letters, invoices and spreadsheets, with a person approving where it matters
- Integrations that act
- The agent connected to a CRM, ticketing, a calendar and GitHub. Rules apply before the model, and every action has a dry run and can be retried safely
- Running on 2 providers
- Anthropic and OpenAI behind one layer, each task routed to one of them, with a fallback when one is down
- Evals and production
- Tests and evals on every pull request, tracing and cost per customer, roles and permissions, and defenses against prompt injection
- Architecture and deployment
- A design where a new customer is configuration, not code. Queues and events. Deployment to a managed platform and to a major cloud with Terraform. Isolation for regulated customers, and delivery into a customer's own cloud account from a runbook
- Research, growth and analytics
- User research inside a deployment, a site that answer engines can cite, outreach within the rules, and an analytics stack with an agent that shows its SQL
What you build
WorkBot is an AI assistant and workflow engine for a company's operations, finance and support teams. It also serves that company's own customers. Employees talk to it in a web app, in Slack and by voice. Customers reach it on WhatsApp, SMS and iMessage.
WorkBot answers from the company's knowledge base and policies, and cites its sources. It takes in work from email, documents and tickets, and applies the company's rules. It asks a person to approve what matters. It acts in the systems the company already runs: the CRM for customer records, ticketing, and the ERP for finance and operations. An admin console lets each customer set it up without code. It hands off to a person when it should.
You build it from the first day and add to it every day.
WorkBot
Operations, finance and support teams inside a company, and the customers they serve
Runs a company's work across its systems. A chat and voice assistant answers from the knowledge base and policies, with citations. It takes in email, documents and tickets, applies fixed rules, asks for approvals and acts in the CRM, ticketing and ERP. An admin console configures it, and it hands off to a person when it should
Teaches: Chat and voice interfaces, where to refuse, images and documents as input, memory, many channels, integrations, roles and permissions, keeping each customer's data apart, workflows that survive a restart, a new customer set up by configuration, personal data under privacy rules, deployment into a customer's own cloud
Features you build along the way, for example:
- A chat assistant that answers an employee's question from company policies and documents, and shows where the answer came from
- A support phone line that listens, speaks back and opens a ticket from the call
- An invoice read from a PDF or a photo, its fields extracted and routed for approval
- An agent that investigates a ticket. It searches past tickets and the knowledge base, checks the CRM and drafts the reply
- Approval workflows that wait for a person, survive a server restart and notify when done
- One conversation that starts on the web and continues on WhatsApp or Slack
- An admin console where a customer sets its own rules, so the next customer is set up with configuration alone
- Dashboards and forms the model composes from a description
- The whole system deployed into a fresh cloud account from a script, with a runbook someone else can follow
- An analytics assistant that answers business questions and shows the SQL it ran
Checkpoints
Every day ends with features merged into WorkBot, and each day is verified from your repository before the next. 3 days are checkpoints: the midpoint design, the delivery into a customer account, and the final release.
- Day 20Architecture design and product thinkingAn architecture document of 2 to 3 pages, a one-page feature proposal, and a 5-minute pitch with questions.
- Day 37User research, and ship to a customer cloudSomeone else deploys WorkBot into a fresh account from the runbook and runs the smoke evals. Interview and usability findings are written up with evidence and routed into configuration and the backlog.
- Day 40Final releaseWorkBot live, an architecture document, a recorded 10-minute demonstration, and the numbers.
Before Day 1
Set up in advance
Needed on Day 1.
- Claude Code: The agentic coding tool used from Day 1, every day of the program.
- Ghostty: Terminal with tabs, splits and desktop notifications, for running several agents side by side.
- Cursor: The IDE, configured for AI-assisted TypeScript.
- GitHub: Repositories and pull requests; the GitHub MCP from Day 3.
- Vercel: Preview and production deployments on every push.
- Codex plugin for Claude Code: Standard and adversarial PR reviews.
- Node.js LTS, pnpm and Git: the base toolchain.
Syllabus
| Day | Title | You ship |
|---|---|---|
| 1W1 | Dev environment and Claude Code fundamentals | A Next.js project scaffolded with Claude Code, and the PR loop run to 10 an hour |
| 2W1 | Claude Code: skills and subagents | 3 custom skills, a pull request review by 3 subagents, and a formatting hook |
| 3W1 | The AI SDLC: feature development workflow | A Linear backlog, with its first tickets shipped through the full loop toward 10 pull requests an hour |
| 4W1 | Multi-project and multi-feature workflow | 4 or more features shipped in parallel, the public site as a second app, and your throughput measured |
| 5W1 | MCP servers and complete feature development | Requests in Postgres, analytics on the intake funnel, and a database MCP server and a PostHog MCP server, both used from Claude Code |
| 6W2 | Anatomy of an AI app | WorkBot's conversation core: a streaming CLI chatbot on either provider, with model switching and cost tracking |
| 7W2 | Tool use and chaining calls | 4 tools running on both providers, and a research command with citations |
| 8W2 | Full-stack AI chat app | WorkBot's web chat deployed on either provider, with tool status and results in the UI |
| 9W2 | Generative UI | An AI dashboard generator with a catalog of 15 or more components, and 3 demo scenarios |
| 10W2 | Context and memory | WorkBot's web chat with context management and memory, deployed |
| 11W3 | AI agent fundamentals | A ReAct investigation agent built from scratch, which shows its reasoning as it works |
| 12W3 | Multi-agent systems and frameworks | A 4-agent pipeline for customer replies, with guardrails, handoffs and tracing, and an intake workflow |
| 13W3 | Modern RAG: agentic retrieval | Agentic retrieval over a 52-document corpus, compared with top-5 retrieval on 10 questions |
| 14W3 | AI scraping, browser use and desktop use | An AI scraper with pagination, and a browser agent that completes a task in a legacy portal |
| 15W3 | The research analyst: orchestration, retrieval and the browser | The AI research analyst, with error recovery, logging, cost tracking and tests, filing its report on the ticket |
| 16W4 | Voice and video AI | A browser voice assistant and a timestamped video analysis tool |
| 17W4 | Structured output and data extraction | An extraction engine with 4 schemas, measured on 18 documents |
| 18W4 | Workflow orchestration with Inngest | A durable document pipeline with human approval, deployed on Vercel |
| 19W4 | Testing, monitoring and deployment | WorkBot made production-ready, with traces, versioned prompts, evals and alerts |
| 20W4 | Architecture design and product thinking | An architecture document, a feature proposal and a 5-minute pitch |
| 21W5 | Multi-CLI development and context architecture | An agentic development setup across 3 CLIs, a company brain, and a one-page guide |
| 22W5 | Agent fleets: 50 concurrent agents, cloud agents and test automation | A 50-task fleet run measured on merges per hour, reverts and brief-to-merge time |
| 23W5 | Agent frameworks: LangGraph.js, Claude Agent SDK, OpenAI Agents SDK | The Day 11 research agent on 3 frameworks, measured on 5 topics, with a one-page comparison |
| 24W5 | 2 providers, 1 codebase | WorkBot's web chat on OpenAI and Anthropic with a routing table, fallbacks and eval scores for both |
| 25W5 | Evals as the development loop | WorkBot on the chosen framework and 2 providers, developed with the fleet workflow, gated by evals in CI |
| 26W6 | Architecture principles: what to unlearn | An architecture audit, 2 migrations measured before and after, and a decision record |
| 27W6 | Integration platforms: Composio, Nango and MCP | WorkBot connected to its ticketing system through the vendor's MCP server, with a synced copy, product-level tools and the 3 routes compared |
| 28W6 | Integration techniques: sync, unified schemas, rules and caching | A unified, rule-applied, cached data layer under the integrations, with reconciliation and an audit trail |
| 29W6 | Config-first design, grammars and meta-compilers | WorkBot config-first, with a second tenant from configuration, a rule language with 20 tests, an entity generator, and its check as a cached task graph |
| 30W6 | Agent security: RBAC, keys and data separation | WorkBot integrated, config-first and hardened, with an injection eval set whose 20 attempts all fail |
| 31W7 | Documents and spreadsheets: PDF, DOCX, XLSX and large tables | A document intake service for PDF, Word and Excel files, with accuracy and cost measured for each path to the model |
| 32W7 | Multi-channel chat: web, SMS, WhatsApp, iMessage and Slack | WorkBot's assistant on 4 channels (web, SMS, WhatsApp and Slack) with shared conversations, delivery callbacks and retries |
| 33W7 | Process lifecycle: queues, pub/sub, SSE, WebSockets and serverless | A processing layer with a fair queue, dead letters, versioned events and streams across 2 instances, load tested with 500 jobs |
| 34W7 | Cloud services: AWS, GCP, Azure and the neo-clouds | WorkBot as two container images on a stack shaped like a major cloud, with deployment steps for a neo-cloud and a major cloud, cost estimates and a recommendation |
| 35W7 | Infrastructure as code: Terraform and deployment targets | WorkBot's infrastructure as Terraform on 3 targets, through a plan, review and apply pipeline, with a demonstrated rollback |
| 36W8 | Isolation, keys and scale thinking | A hardened, optionally single-tenant deployment with runbooks, a game day and a completed security questionnaire |
| 37W8 | User research, and ship to a customer cloud | A findings page from interviews and usability sessions, and a customer-style deployment of WorkBot proven by someone else from the runbook |
| 38W8 | AEO and growth automation | WorkBot's site passing an AEO checklist in CI, a 200-account target list, and a compliant outreach sequence |
| 39W8 | Research and analytics with agents | A published research dossier, the analytics stack with a semantic layer, and a read-only analytics agent |
| 40W8 | Final release | WorkBot live, an architecture document, a recorded 10-minute demonstration and the numbers |
Questions
- What is the AI FDE Course?
- Free forward deployed engineer (FDE) training from Plank: 40 days in which a full-stack engineer builds and ships one production AI product, WorkBot, the way an FDE does. It is the Plank AI Accelerator's 2-month program, open to anyone.
- Who is the course for?
- Full-stack engineers who work in TypeScript and Node.js and want to move into AI engineering or a forward deployed engineer role, at any stage of their career. No prior AI engineering experience is required.
- How long does the FDE training take?
- 40 days of work, in 8 weeks or faster. Week 1 sets up the AI-powered development workflow; weeks 2 to 8 build WorkBot through chat, agents, retrieval, voice, evals, integrations, security, infrastructure and a final release.
- Is the course free?
- Yes. Every day, the verification and the profile are free. You pay for your own model and cloud usage.
- How is each day verified?
- Sign in with GitHub and link a public repository. When you finish a day, push and verify: the commits since your last verified day are checked against what that day ships, and the day is recorded with its commit.
- What do I have when I finish?
- WorkBot deployed, and a public profile on joinplank.com, under your GitHub login, listing every verified day, its commit and the skills it covers. Plank hires forward deployed engineers, and a verified profile is a record of shipped work you can link in an application.
- How does the course relate to the Plank AI Accelerator?
- The course is the Accelerator's 2-month program, open to anyone and done at your own pace. The full Accelerator is a selective one-year program, run in person, that ends with a role at Plank as a forward deployed engineer.
Anusheel Bhushan co-founded Plank after serving as co-founder and CTO of Plato.ai, an Accel-funded company that raised a Series A. Anusheel has led software product development for more than 80 companies in 10 countries, mentors founders at Stanford StartX and holds an MS in Computer Science from UC Irvine.
Curriculum updated 2026-10-26.