Autonomous Post-Training: Build a 1B In-Car Agent

Build RoutePilot, a 1B-class route-aware agent that extracts constraints, calls a maps or POI tool, ranks returned candidates, and proposes a safe next action. The course starts with a deterministic behavioral contract, then moves through licensed synthetic data, baselines, post-training, failure-driven iteration, structural pruning with recovery training, quantization, and publication.

3 modules~6 hours total

About This Course

A useful in-car language model should not memorize restaurants, opening hours, traffic, or charger availability. Those facts change. RoutePilot learns the stable part of the job: understand the request, convert it into explicit constraints, select a tool, reason over current results, and explain a proposed action.

The course treats post-training as an experimental system. Every model iteration is evaluated against deterministic hard constraints and ranking utilities. Failures become categorized evidence for the next data pass. No accuracy, latency, or deployment claim is made until a reproducible run is committed.

Course design and benchmark are original to SciMigo. Product and model claims are paraphrased from linked primary sources fetched 2026-09-21. Google, Cerence, Quadrillion, Qwen, and Hugging Face are trademarks of their respective owners; this independent course is not endorsed by them.

Prerequisites

  • Comfortable Python and JSON
  • Basic familiarity with language-model prompting and supervised fine-tuning
  • No prior automotive or edge-ML experience required

What You Will Learn

  • Define a testable product contract for a tool-using route assistant
  • Generate licensed training scenarios whose labels follow explicit constraints
  • Post-train and compare a 1B-class student against teacher and base-model baselines
  • Diagnose hard-constraint, tool-call, ranking, and clarification failures separately
  • Compare an adapted student with structurally pruned, recovered, and quantized variants on named edge hardware
  • Publish a model card, data card, benchmark, and integration boundary without overstating results

Terminology Mapping

How classic concepts map to the terminology used in this course.

ClassicThis Course (Routepilot)
Dynamic world knowledgefacts returned at request time by maps, POI, traffic, or charger tools
Hard constrainta condition that disqualifies a candidate, such as closed, wrong connector, or too much detour
Soft preferencea weighted signal that orders feasible candidates, such as rating or lower detour
Oracledeterministic code that applies the declared constraints and utility weights
Agent actiona proposal that still requires application policy and, when appropriate, driver confirmation

Your Learning Path

Each module builds on the last. Take your time—the AI tutor is with you at every step.

1

The Contract Before the ModelFreeze what the agent may know, what it must call, and how a wrong answer fails

Turn an in-car request into an explicit tool contract and a deterministic evaluation. Separate changing world facts from model behavior, hard constraints from soft preferences, and extraction errors from ranking errors before generating a single training example.

90 minReading materialLab exercises
2

Generate Scenarios, Not AnswersCreate structured worlds, derive labels with code, and keep template families out of both sides of a split

Build a deterministic scenario factory for RoutePilot. Generate contexts and candidates first, derive every expected choice with the oracle, separate template families across train and development, and allow language models to vary wording only after structured truth exists.

120 minReading materialLab exercises
3

Measure the BaselinesFreeze the prompt, preserve every raw response, and decide whether training is justified

Run the same RoutePilot development scenarios through an unmodified 1B-class student and a named teacher. Record prompts, raw responses, parsing failures, usage, costs, and component metrics so a later post-training claim compares like with like.

150 minReading materialLab exercises