Naman Pandey Curriculum · Draft v1 · Nov 2026
Forward-deployed engineering

From tokens to agent systems in twelve weeks.

A 12-week program for new grads in CS or AI. They start at "what is a token" and finish by shipping a reliable multi-agent system into a customer's actual systems.

Length12 weeks
Phases4
Builds12 + capstone
DomainOne mock enterprise
Principle

Build every layer from scratch once

They write the raw agent loop before using an SDK, so the frameworks never feel like magic.

Principle

Evals in week 4, not week 12

In FDE work, being able to prove the system works matters more than adding features.

Principle

One company all quarter

Acme Logistics has a CRM, ticketing, Postgres, a wiki, and an old portal with no API. Each week's build adds to it.

Principle

Craft every week

Discovery, scoping, demos and writing run as a weekly thread, not one lecture at the end.

Week 0 prereq (self-serve): Python, git, HTTP/REST, SQL, Docker, one cloud account. Gate: a FastAPI service in Docker that calls a public API.

Phase 1 · Weeks 1–4

How LLMs actually work, then the first app

Week01

LLM foundations: tokens, prediction, generation

BPE tokenization and why "strawberry" is hard · embeddings · the transformer as a next-token predictor · logits, probabilities and sampling (temperature, top-p) · context windows · pretraining vs SFT vs RLHF · why models hallucinate · tokens as the unit of cost and latency.

Build

A token lab notebook. Compare token counts across languages and in code, pull logprobs, and change temperature to watch the output distribution shift.

FDE craft

Explain to an exec in two minutes why the model "made something up."

Week02

Talking to models: APIs and prompting

The Messages API and stateless multi-turn · system prompts · streaming · structured outputs and JSON schemas · clear instructions, examples, XML tags, chain of thought, prefill · prompt caching · picking a model on cost, latency and quality.

Build

A chat assistant with streaming, conversation memory, a structured-output mode and a live token and cost meter.

FDE craft

A one-page model-choice memo: quality, cost per 1K requests, latency.

Week03

First real app: grounding in your own data

Embeddings and vector search · chunking · hybrid search (BM25 + vectors) · reranking · contextual retrieval · citations · when long context beats RAG.

Build

Ask Acme: a deployed web app that answers questions over the company wiki, with citations.

FDE craft

Turn a messy customer data dump into a corpus you can actually search.

Week04

Evals: proving it works

Error analysis before metrics · golden datasets · code checks vs LLM-as-judge · regression tests for prompts · offline vs online evals · tracing.

Build

An eval harness for Ask Acme: 50+ graded cases, and CI that fails on regressions. Use it to justify one change with data.

FDE craft

Turn "make it better" into success criteria you can measure.

Phase gate

A deployed RAG app, an eval report, and a five-minute demo to "the customer."

Phase 2 · Weeks 5–7

Tools and integration: connecting models to real systems

Week05

Tool use and calling APIs

How function calling works (the model asks, your code runs it, the result goes back) · tool schema design · parallel calls · errors, retries, idempotency · tool descriptions as prompts.

Build

An Acme ops assistant that calls the CRM and ticketing REST APIs: look up a customer, open a ticket, check a shipment.

FDE craft

Read a customer's ugly API docs and wrap them cleanly.

Week06

MCP: the standard plug for tools and data

Host, client and server · tools vs resources vs prompts · stdio vs streamable HTTP · OAuth for remote servers · MCP vs a direct API call · tool poisoning and over-broad scopes.

Resources
Build

An Acme MCP server over Postgres, the CRM and the wiki, used from Claude Desktop or Claude Code and from their own client.

FDE craft

Write the security and permissions section of an integration design doc.

Week07

Browser use and computer use

Enterprises run on legacy portals with no API · DOM and accessibility-tree agents vs screenshot agents · Playwright as the base · flaky selectors, auth, waits · prompt injection from web content · sandboxing.

Resources
Build

An agent that finishes a multi-step workflow on Acme's legacy portal (a sandboxed site), logging a screenshot at every step.

FDE craft

Handle the security conversation with the customer's IT and InfoSec.

Phase gate

Integration day: connect the system to a mock system nobody has seen before, in one day.

Phase 3 · Weeks 8–10

Agents: from toy to semi-autonomous

Week08

The first toy agent, from scratch

An agent is an LLM using tools in a loop · workflows vs agents · prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer · stopping conditions · ReAct.

Resources
Build

A ~150-line agent with no framework: loop, tool registry, step limit, logs. Then rebuild it on the Agent SDK and compare.

FDE craft

Decide whether the ask needs an agent at all, and defend the answer.

Week09

Context engineering and memory

Context as a finite budget · compaction and summaries · scratchpads and progress files · just-in-time retrieval · short- and long-term memory · skills and instruction files · subagents to keep context clean.

Build

Make the agent survive a one-hour task (reconcile 500 Acme invoices) using compaction and a progress file.

FDE craft

Demos that build trust, including how to fail gracefully in front of the customer.

Week10

Semi-autonomous: humans in the loop

Approval gates on side effects · interrupt and resume · durable state and checkpoints · read / draft / execute permission tiers · escalation · audit logs.

Build

An Acme collections agent that triages overdue accounts, drafts outreach, pauses for approval, and resumes after a restart.

FDE craft

Agree with the customer on which actions need human sign-off.

Phase gate

A live demo while a reviewer tries to break it: bad inputs, prompt injection, killing the process mid-task.

Phase 4 · Weeks 11–12

Multi-agent systems and enterprise readiness

Week11

Multi-agent systems and handoffs

Orchestrator-worker vs peer handoffs vs pipelines · what passes at a handoff · shared state · A2A agent cards and the task lifecycle between agents, with MCP for tools inside each · token cost of multi-agent setups · when not to go multi-agent.

Build

Order-to-resolution: an intake agent hands off to billing, logistics (browser) and CRM (MCP) specialists, which report back to a supervisor. Traced end to end.

FDE craft

An architecture tradeoff memo: single agent vs multi-agent, with cost.

Week12

Production, security, and the capstone engagement

OpenTelemetry tracing · cost controls and rate limits · secrets and agent identity · PII and data residency · OWASP LLM Top 10 · red-teaming · deploying to the customer's cloud (Bedrock, Vertex, Azure) · driving adoption.

Capstone

A simulated engagement. An instructor plays a customer with a vague brief; engineers run discovery, a scoping doc, the build, evals, a red-team report, a live demo and a handover doc.

FDE craft

Discovery interviews, SOWs, handover and driving adoption.

The craft thread

30 to 60 minutes every week on the part of the job that isn't code.

W01–02Explaining LLM behavior to non-technical people
W03Getting access to customer data; triaging messy data
W04Turning "make it better" into measurable success criteria
W05–06Reading hostile docs; integration design docs
W07Security conversations with IT and InfoSec
W08"Do you even need an agent?" scoping
W09–10Demos that build trust; handling failure live
W11Architecture tradeoff memos
W12Discovery interviews, SOWs, handover, adoption