QandA_AgenticAI_TrainingRFI.xlsx
XLSX spreadsheet 24 KB Posted
- Solicitation number
- 6HQOIG-26-A-0003
- Issued by
-
United States Postal Service
About this file
This is a Questions and Answers (Q&A) document addressing an RFI for Agentic AI and Advanced GenAI Training Support Services for the United States Postal Service (USPS) Office of Inspector General (OIG). The USPS OIG's Regulatory Innovation and Strategy Center (RISC) seeks to develop training curricula and conduct hands-on lab instruction for approximately 40–50 employees across two cohorts, with spring/summer 2024 as the target delivery window.
The training will occur on Government-Furnished Equipment (GFEs) running Windows OS with VSCode, Databricks, Microsoft AI Foundry, and access to Google Gemini, NotebookLM, and AskSage LLM APIs. The government will provide an Azure VM lab environment with agreed-upon libraries; vendors must coordinate pip install requests through IT Security. All participants have Python proficiency, though some prefer SQL; very few have experience with agentic frameworks. The curriculum should introduce LangChain/LangGraph, CrewAI, and AutoGen frameworks—none currently in production—to help RISC evaluate and select a standard. Training will emphasize synthetic contractor-provided data for RAG labs rather than internal datasets, with key outcome metrics including retrieval and accuracy (RAG Triad), logging, monitoring, and regulatory compliance. Delivery is in-person at Arlington, VA (default) with virtual Teams option; the USPS OIG will provide classroom space, A/V, and facility hosting. Instructors need only standard front-desk check-in with no citizenship or clearance requirements. The government retains unlimited data rights to instructional and training materials per FAR Clause 8-6 Data Rights, while the vendor retains rights to its developed synthetic data.
View the file
Other files
| File | Type | Posted |
|---|---|---|
| Draft Statement of Work - Agentic AI.pdf | ||
| RFI - Agenic AI Training Support - SAM.pdf |
On GovTribe
Work with this file
- Download the original file
- Contacts named in this file
- Similar government files
- Ask GovTribe AI about this file
Text version
Sheet1
| No. | Reference | Question | Answer | |
| 1 | Technical Environment & Tooling: To ensure the "Vibe Coding" modules are actionable, what development environment will students utilize during training? Specifically, will analysts have access to a cloud-based sandbox (e.g., AWS SageMaker, Azure AI Studio) with unrestricted pip install privileges, or will training occur on Government Furnished Equipment (GFE) laptops? | |||
| • | Context: Understanding this constraint allows us to determine if we can utilize cloud-native IDEs like Google Antigravity or if we must tailor the curriculum for local VS Code environments. | Training will occur on GFEs. We are able to use Databricks, Microsoft AI Foundry, and VSCode. We are able to pip install but will need the vendor to let us know what they want us to install so that we can get it cleared through ITSec first. |
We have access to Google Gemini and NotebookLM (but not Google API to use Gemini models programmatically). We are working on getting set up with Google CLI and Github Copilot.
We also have access to AskSage (including API) for additional LLMs.
| 2 | LLM Access & API Constraints: Does the RISC team currently possess an Authority to Operate (ATO) for specific external Model APIs (e.g., Azure OpenAI GPT-4o, Anthropic Claude via Bedrock), or does the agency prefer the curriculum to focus on open-weights models (e.g., Llama 3, Mixtral) running locally via tools like Ollama? | See answer number 1. |
| 3 | Participant Baseline Proficiency: To calibrate the "Agentic Frameworks" curriculum, could you provide the approximate technical breakdown of the expected cohort? Specifically, what percentage of the audience is currently proficient in Python scripting versus those primarily utilizing SQL/Excel/Tableau? | All participants have proficiency in python but some are more inclined towards utilizing SQL. Very few have experience with agentic frameworks and utilize GenAI for smaller tasks. |
| 4 | RAG Data Sources: For the "Advanced Prompt & Context Engineering" requirement, is the agency's preference to conduct RAG (Retrieval-Augmented Generation) labs using synthetic, contractor-provided data, or does the OIG intend to test these agents against actual, sanitized internal datasets during the course? | We would prefer to use synthetic contractor-provided data instead of using our own internal data for the training. |
| 5 | Framework Standardization: The SOW mentions LangGraph, CrewAI, and AutoGen. Does the USPS OIG have an existing preferred framework for production deployment, or is the goal of this training to help the RISC team evaluaand select a standard framework? | We have built prototypes in LangChain / LangGraph, CrewAI, and AutoGen but none are in production. Most participants will not have experiemce in any of those so it is good to introduce all of them so that RIS can further evaluate. |
| 6 | Training cadence/volume: How many cohorts/classes do you anticipate (cap 20 students per class), and is the intent a single delivery or multiple deliveries over the period of performance? | We believe we will have 2 classes on different dates so that students can determine better timing. |
| 7 | Target timing: Do you have an estimated target month/week for the first delivery (and any follow-on sessions)? | Spring/Summer |
| 8 | Delivery modality: Should we assume in-person (Arlington, VA) is the default, or will virtual delivery be acceptable/preferred for some sessions? | In person is the default with option of virtual |
| 9 | Virtual platform: If virtual, is there a required platform (Teams/GovZoom/etc.) and will the Government provide hosting/accounts? | Teams - we are happy to create the meeting and invite the trainers to the meeting. |
| 10 | Hands-on lab environment: Will participants use Government-furnished equipment/network, or should the vendor provide a lab/sandbox? Any restrictions on internet access, external GitHub, commercial model APIs, or cloud sandboxes during labs? | We will provide an Azure VM (running a Windows OS) with VSCode and agreed-upon libraries. This VM will be in a LAB environment and cannot contain any internal or sensitive data but may use publicly-available or synthetic-data. We can also provide synthetic datasets. |
| 11 | Security/vetting: Are there any instructor requirements for citizenship, background checks, public trust/clearance, or onsite badging? | Trainers will need to check in at the front desk but no other requirement |
| 12 | Facilities: For in-person training, will USPS OIG provide classroom space/AV, or should the vendor plan to provide space and A/V? | USPS OIG will provide the space/AV. |
| 13 | IP/data rights: Are there any expectations for Government rights to training materials/labs or restrictions on reuse of vendor-developed content? | Unlimited Data Rights for any "Instructional and Training Material" delivered IAW Clause 8-6 Data Rights. |
We retain all rights to any synthetic data that we provide to you that we developed.
| 14 | Have you identified specific analyst or audit workflows (e.g., data pulls, brief drafting, evidence triage, report QA) where agentic automations or vibe‑coding accelerators are most likely to close current pain points? If so, can you share priority use cases and constraints (data access, tools, policies) to anchor course and lab design? | We are looking to develop more webapps through our data to deliver data insights. Currently, we are developing in PBI and want to expand to other forms of development for analytics products. |
| 15 | What are the IT/security requirements and compute constraints for hands‑on labs (e.g., connectivity, model access—commercial vs open‑source, offline options)? | We will provide an Azure VM (running a Windows OS) with VSCode and agreed-upon libraries. This VM will be in a LAB environment and cannot contain any internal or sensitive data but may use publicly-available or synthetic-data. We can also provide synthetic datasets. We pay for compute in this space so we want to be cognizant of costs. |
| 16 | Will the team receive a breakdown of expected attendee roles (analysts vs DS/DE vs leadership) and any baseline skills assessment to tailor tracks and capstone labs accordingly? | We can break down the attendee roles to help tailor the labs accordingly. |
| 17 | Which training outcome metrics matter most for the program (e.g., knowledge gains, time‑to‑prototype agent, defect rate in AI‑generated code, retrieval precision/recall in RAG, policy compliance)? | 1. Retrieval & Accuracy Metrics (The "RAG Triad") |
2. Logging and Monitoring
3. Scope, Law, and Regulation Compliance
| 18 | Can USPS OIG please confirm that the training audience is limited to RISC and what is the total size of the audience? | It will be limited to RISC and the total size of the component is 50 individuals (not everyone will take the training) |
| 19 | Would USPS OIG please provide detailed information regarding the current knowledge and capabilities of the training audience described in the RFI? | Our analysts and data scientists have experience using Databricks and coding languages within it. We have a handful of analysts and data scientists that are developing in VSCode and are using GenAI for coding assistance. |
We have a handful of advanced data scientists that are familiar with LLMs and RAG.
Our research team is less technical and relies more on GenAI for report synthesization, research, and writing. They have minimal coding experience (if any).
| 20 | Can USPS OIG please confirm that the Cover Letter is considered outside of the overall page limit? | Correct. |
| 21 | Can USPS OIG confirm this training will leverage the USPS training platform (LMS) in any capacity? | No |
| 22 | Does USPS OIG envision one training curriculum targeted toward the overall audience or multiple training curriculums tailored to target audiences (e.g., data analysts, data scientists, data specialists, researchers, managers, directors, and executives)? | We would like an overall curriculum of understanding concepts and the art of the possible. When we move into the lab environment, we will most likely only provide that for our data analysts and scientists. People that have coding experience. |
| 23 | Which agentic-enabled technologies has USPS already adopted within the enterprise? | Microsoft CoPilot Studio, Gemini Pro, AskSage, Jira/Confluence (Rovo) |
| 24 | How is access and credentialling intended to be handled for agents? Carry user permission? Separate for agents? Hybrid? | Not finalized. Most likely, many agents will be user-delegated permissions based on Microsoft Entra (e.g., CoPIlot Studio) but some might use managed identity / service principals as needed. |
| 25 | What Workplace training platforms does USPS currently use? | DataCamp, LinkedIn Learning |
| 26 | Approximately how many employees will be trained in person? | 40-50, over period of two courses. |
File details come from the government source that posted it. Updated .