Title: Module 2 - Ship (Evals, Guardrails, Economics & Production) Syllabus - Google Reichman AI Tech School

URL Source: https://graitechschool.co.il/module-2-ship-evals-guardrails-economics-production-syllabus

Markdown Content:
*changes may apply

![Image 1](https://www.graitechschool.co.il/wp-content/uploads/2023/02/Group-11870.png)

![Image 2](https://www.graitechschool.co.il/wp-content/uploads/2022/12/%D7%A0%D7%99%D7%A0%D7%94-%D7%A2%D7%A0%D7%AA-%D7%95%D7%90%D7%93%D7%99%D7%A1%D7%9C%D7%90%D7%9D-scaled.jpg)

Program Overview

This module transitions teams from fragile sandboxes to hardened, dependable production deployments. Built with an agentic-by-default focus, it teaches teams how to run automated evaluation suites, deploy runtime safety filters, and optimize multi-step token unit economics.

![Image 3](https://www.graitechschool.co.il/wp-content/uploads/2022/12/Gilad-Kavalerchik__221220_130.jpg)

Session 1: From Prototype to Production — Metrics That Matter

How do you know an AI product — including an agentic one — is actually working, and tie that to release decisions?

 Core Hands-on Competencies

*   Strategic AI Telemetry: Designing multi-layered performance metrics: tracking leading and lagging usage indicators, task completion velocities, and user behavioral corrections (e.g., acceptance, edit, retry, and system escalation rates).
*   Operational Task Economics: Establishing cost-per-successful-task calculations and manual human-review rates as first-class, baseline product health metrics.
*   Probabilistic Quality Frameworks: Constructing scoring definitions for non-binary software behaviors, including system precision, recall accuracy, context faithfulness, groundedness, and semantic relevance. Setting explicit confidence thresholds to manage trade-offs between false positives and false negatives.
*   Agentic Trajectory Mapping: Measuring autonomous multi-step execution health by calculating trajectory success rates, step execution efficiency, tool-call destination accuracy, and token-cost-per-completed-task metrics.
*   Automated Release Gating: Mapping explicit quality criteria thresholds directly to automated go/no-go product deployment rules.

Additional topics that will be shortly reviewed:

*   The Operator’s Framing: Navigating the technical and organizational shift required to transition a fragile sandbox demo into a dependable, hardened production-grade software system (30 min).

Hands-on Workshop: Participants build a comprehensive, multi-layer metrics dashboard tree for their product feature, mapping at least one autonomous agentic trajectory health path and a hard, metric-driven production release gate.

 Session Deliverable: AI Metrics & Release-Gate Plan.

![Image 4](https://www.graitechschool.co.il/wp-content/uploads/2024/02/Gilad-Kavalerchik__221220_139-scaled-e1708860677514.jpg)

Session 2: Evals & Golden Datasets

What does “good enough to ship” mean, and how do you test a probabilistic, multi-step system?

 Core Hands-on Competencies

*   Golden Dataset Engineering: Curating and formatting highly representative evaluation datasets, including defining system prompt inputs, expected outputs, explicitly unacceptable output responses, and precise alignment rubrics.
*   LLM-as-Judge Automation: Configuring and tuning automated LLM-as-judge prompts, establishing scoring constraints, and calibrating the automated evaluator against human grading baselines to avoid judge bias.
*   CI/CD Regression Deployment: Deploying automated evaluation runs into software delivery pipelines to catch prompt regressions and system drift before push commands execute (using Promptfoo and Braintrust).
*   RAG Pipeline Evaluation: Implementing testing matrices to evaluate RAG configurations, specifically testing context retrieval quality, source material faithfulness, and hallucination containment.
*   Agentic Trajectory Evals: Designing multi-step evaluation checkpoints that audit and score the entire operational path, tracking intermediate tool-use correctness rather than simply checking the final terminal answer string.
*   Compliance & Tone Validation: Structuring automated semantic tests to verify corporate tone alignment, brand safety guidelines, and regulatory compliance constraints.

Additional topics that will be shortly reviewed:

*   Understanding why traditional deterministic software QA models fail completely when testing non-deterministic, probabilistic AI behaviors.

Hands-on Workshop: Participants assemble a functional golden dataset and grading rubric, execute a live automated prompt evaluation using the Promptfoo CLI tool, and configure an automated multi-step trajectory path validation test.

 Session Deliverable: Eval Suite — Golden Dataset, Evaluation Rubric, and Automated CI/CD Testing Gate.

![Image 5](https://www.graitechschool.co.il/wp-content/uploads/2022/12/Gilad-Kavalerchik__221220_174-1-e1671950437337.jpg)

Session 3: Guardrails, Agent Safety & Monitoring

How do you keep an AI product — and an autonomous agent — safe, governed, and observable in production?

 Core Hands-on Competencies

*   Live Guardrail Filtering: Setting up input and output filters designed to block prompt injections, intercept PII leaks, filter toxic content, stop unsafe system recommendations, catch policy hallucinations, and prevent unauthorized downstream actions.
*   Agentic Circuit Breakers: Coding structural system boundaries for autonomous agents, implementing strict execution step caps, token budget caps, systemic loop detection routines, narrow tool-scoping parameters, and human-approval gates on high-risk actions.
*   Action Authorization Matrix: Designing a strict permissions architecture specifying exactly what tasks an agent can execute completely autonomously vs. which tasks must halt execution to request explicit user approval.
*   Data Fencing Schemas: Creating architectural boundary maps that govern exactly which corporate data assets are permitted to flow to specific downstream models, external vendors, or processing endpoints.
*   Auditable Compliance Logging: Designing tamper-resistant decision and transaction history logs structured to survive formal internal or external regulatory reviews.
*   Live Production Telemetry: Deploying real-time monitoring post-launch to detect semantic data drift, gradual model performance degradation, vendor update regression risks, and step-level transaction tracing for autonomous agents.

Additional topics that will be shortly reviewed:

*   Treating system risk as a core product requirement, exploring Safety-by-Design and Zero-Trust infrastructure frameworks, and reviewing localized compliance mandates (such as the EU AI Act risk classifications and reporting obligations).

Hands-on Workshop: Participants map out an end-to-end runtime security framework, producing a functional guardrail and agent circuit-breaker checklist, a data-fencing flow schema, and a production alerting and monitoring plan.

 Session Deliverable: Guardrails, Agent-Safety & Live Monitoring Plan.

![Image 6](https://www.graitechschool.co.il/wp-content/uploads/2022/12/Gilad-Kavalerchik__221220_58-e1671953361203.jpg)

Session 4: Model Selection, Unit Economics, Build vs. Buy & Production Readiness

How do PMs make the strategic calls on models, cost and margin, and what it actually takes to ship?

 Core Hands-on Competencies

*   Model Optimization Architecture: Evaluating trade-offs across commercial APIs vs. open-source weights, RAG architectures vs. fine-tuning, small task-specific models (SLMs) vs. large general foundation models (LLMs), latency-cost-quality optimization vectors, context windows, and dynamic model routing tables.
*   Granular Cost-Bucket Modeling: Deconstructing the traditional COGS equation down into discrete Input, Output, and Retrieval cost buckets. Applying targeted engineering levers to maximize margins (e.g., prompt caching optimization to reduce input costs, strict max-token caps to minimize output costs, and semantic chunking to decrease retrieval costs).
*   Agentic Financial Control: Building token multiplication projection models for multi-step agent behaviors, linking application cost paths directly back to the execution circuit-breaker caps.
*   The 3× Value Heuristic: Applying financial stress testing to verify that an AI capability generates measurable business value at least three times greater than its direct compute cost, customizing the model against company gross margin targets and Customer Acquisition Cost (CAC) structures.
*   Pricing & TCO Calculations: Building robust Total Cost of Ownership (TCO) matrices spanning costs per request, per active user, and per resolution; analyzing pricing frameworks across usage-based, subscription, hybrid, or internal corporate chargebacks.
*   Cross-Functional AI Governance: Constructing a specialized RACI matrix mapping operational ownership across Product Management, Engineering, Data Science, Security, and Legal teams for long-term model maintenance, evals, guardrail logic, and cost allocation.

Additional topics that will be shortly reviewed:

*   Build vs. Buy evaluation criteria, filtering marketing claims, managing enterprise AI

 procurement processes, navigating security and compliance reviews, and reviewing engineering delivery workflows (CI/CD pipelines, code testing, and auditing AI-generated production code).

Hands-on Workshop: Participants construct a live, parameter-driven TCO and COGS spreadsheet model for their feature, run the 3× financial value check, map out a lifecycle RACI governance framework, and compile a final deployment readiness plan.

 Session Deliverable: AI Product Production Plan + Executive Decision Brief (comprising the detailed COGS breakdown, 3× value validation, and cross-functional lifecycle RACI).

What Participants Graduate With

Participants exit this program equipped with a functional portfolio of production-ready blueprints. They will know how to utilize Claude and Claude Code to accelerate discovery, generate machine-readable PRDs, and construct functional interface prototypes.

 More importantly, they will leave possessing the core competencies needed to confidently handle the complex, hidden vectors that determine whether an enterprise AI application can scale sustainably: establishing rigorous evaluation criteria, configuring safety guardrails, monitoring multi-step agent trajectories, and structuring optimized unit economic models to guarantee long-term profitability.