The Last Mile wanted to teach AI literacy to incarcerated students — not a demo, but real access to the frontier models reshaping the workforce. The problem: no platform existed that could satisfy the security and compliance requirements of a correctional facility. TDG built one from scratch. And once the secure infrastructure was in place, it unlocked something bigger: a modular simulator platform that can bring any digital workplace tool — email, spreadsheets, coding environments, AI prompting — safely into classrooms that have never had access to any of it.
The Last Mile runs award-winning technology education programs inside correctional facilities across the United States. When they decided to add AI literacy to their curriculum, they ran into a problem no one had solved.
Any platform hoping to bring modern, intelligent tools into a prison classroom has to do more than solve a technical problem. It has to earn institutional trust — and that means transparency, auditability, and documented proof that every precaution is being taken.
The security challenge had three distinct layers. First: create an impenetrable barrier between the student and the model — an LLM with even a strong system prompt is still vulnerable to hallucination and drift, and a single failure in a correctional setting ends the program. Second: even a perfectly secure system isn't enough without full transparency — DOCs need to see the record, not just trust the result. Third: if the platform was going to scale across an entire curriculum, security couldn't live inside each simulator. It had to be centralized — one place to control, one place to audit, one architecture that every future simulator could inherit.
What TLM needed wasn't a safer wrapper around an LLM. They needed a centralized platform that could prove its own safety — and once that platform existed, it unlocked something much bigger: the ability to bring the entire digital workplace into the classroom for the first time.
TDG built the architecture from first principles — designing a system whose security properties come from its structure, not from trusting any single component to behave correctly. Once the secure brain was in place, the curriculum opportunity became clear: any workplace tool could be simulated through the same pipeline, with the same intelligent, rubric-scored feedback.
All student interactions route through a single backend before any AI call is made. Security isn't distributed across simulators — it lives in one place, is controlled in one place, and is logged in one place. Adding a new simulator means conforming to a standard payload contract. The security infrastructure handles the rest automatically.
Every student query clears three sequential filters before reaching the model: a deterministic regex blocklist, a hardcoded common-sense contraband prompt, and a dynamic keyword list that TLM administrators can edit in real time — no code deployment required. No unfiltered content reaches the LLM. Security is structural.
Evaluation runs first. A student's score and rubric feedback are preserved before the LLM response is generated — so even if a response is flagged on its return trip through the sanitization layer, the student still receives their assessment. Learning continues regardless of content outcomes.
Every transaction — student ID, facility, simulator, score, flag status, timestamp — is logged and surfaced in a filterable admin dashboard. This is what earns DOC approval: not a promise of safety, but a verifiable, ongoing record of it.
Any learning experience that can be designed is deployable through the same secure pipeline. TDG built 20 simulators to launch the AI literacy curriculum — a Gmail-style inbox, a Prompt Lab, an AI literacy module, and more — each delivering intelligent, rubric-scored feedback that sends students back to try again until they get it right. TDG then transferred the build process itself to TLM's curriculum team, including a purpose-built tool that auto-generates and encrypts scoring rubrics. The team can build new simulators without learning prompt engineering.
TDG funded the build and handed TLM a production-ready platform with full ownership — the repos, the simulators, the factory process, and an internal team trained to carry it forward. The investment was in the sector. The asset belongs to the mission.
Someone incarcerated even twelve months ago will be coming home to a world radically transformed by AI. This platform exists so that transformation isn't a blindside.
The first cohort launches at San Quentin in May 2026 — 11 students working through a full AI literacy curriculum on a platform that gives them real access to frontier models, real rubric-scored feedback, and real workplace tools. A second cohort follows in Indiana in June.
The simulators look exactly like the software students will encounter on the job. The Gmail simulator scores email responses against professional rubrics and sends the student back to try again until they pass. The Prompt Lab teaches students to write effective AI prompts and scores their work against explicit criteria. Every simulator runs through the same secure, auditable brain.
Someone incarcerated even twelve months ago will be coming home to a world radically transformed by AI — where the tools that define modern hiring, modern communication, and modern work have all shifted. This platform exists so that transformation isn't a blindside.
A full AI literacy curriculum deployed across 20 purpose-built learning environments — email, prompt labs, AI literacy modules, and more — all running through a single secure backend with intelligent, rubric-scored feedback.
First cohort of 11 students. The first AI literacy program delivered inside a correctional facility with real frontier model access, intelligent assessment, and a full compliance record.
Triple-layer sanitization built from first principles for environments where a single compliance failure ends the program. Security lives in the architecture, not in model behavior.
TDG funded the build and handed TLM a complete platform — repos, simulators, build process, and trained internal team. The investment was in the sector. The asset belongs to the mission.