Lab.IA
An AI solutions factory designed for a state government. Architecture delivered. Implementation in progress.
A state government entered 2026 with more than 15 AI initiatives running across at least 10 secretariats, two active innovation labs, and installed GPU infrastructure. The vocation was there. The structure was not.
Each initiative had been built independently. No shared platform. No common governance. No visibility into cost or model use. Whenever a new use case emerged — a copilot for the budget office, an audit agent for the comptroller, a triage system for health — it started from scratch. Same integrations rebuilt. Same data wrangling re-done. Same costs paid twice.
The state’s digital-transformation secretariat had read what public-sector AI benchmarking across fifteen countries showed. Governments that succeed at scale build shared platforms with centralized governance and incremental adoption. They came to Quantum Leap to architect one for the state from scratch.
What was at stake: regulatory compliance under the AI Act and national AI regulation, fiscal accountability with hard auditability, and a window of executive support that demanded the first piloto in production within four months.
The engagement folds three horizons into a single architecture.
Horizon 01 mapped the existing 15+ initiatives across secretariats, classified them by maturity, and surfaced where a shared platform would create the most leverage. Discovery framed the strategic choice: rebuild what already exists (politically and technically impossible) or design a platform that becomes the standard for new solutions while leaving existing initiatives to operate. The second path was taken.
Horizon 02 designed the platform's intelligence layer. A unified LLM Gateway routes to GPT-5.4 Thinking with failover. OpenAI File Search handles document retrieval for Waves 0–1, with a proprietary RAG planned for Waves 3–4. The audit and governance stack was designed alongside, not bolted on.
Horizon 03 designed the agent layer that automates Quantum Leap's own delivery pipeline and the secretariat-facing agent primitives that compose into new solutions. A 9-stage delivery process (Intake → Discovery → Analysis → Design → PoC → Build → QA → Go-Live → Operation) became the operational template.
Implementation runs in partnership with the secretariat’s team. Quantum Leap delivered the architecture and built the first wave of agents. The platform is in active rollout under the Secretariat’s operational control.
A three-layer architecture, designed once and deployed iteratively.
Layer 01 / Secretariat Apps. Each secretariat owns its own application (web or mobile), with its own deployment cycle, consuming the Core via API. Isolation is intentional. No secretariat blocks another.
Layer 02 / Core Platform. The shared intelligence layer. LiteLLM Gateway routing to GPT-5.4 Thinking with failover. OpenAI File Search for document retrieval. Governance designed in from week one: RBAC across 7 roles, end-to-end tracing via Grafana and Langfuse, consumption ledger by project_id with budget alerts, CI/CD pipelines with Trivy, Semgrep, and Gitleaks as security gates.
Layer 03 / Cloud Services. OpenAI APIs, Supabase with multi-tenant RLS, Railway hosting. Cloud-first with direct operational control. No third-party datacenter dependencies. No provisioning queues.
The agent layer operates on two fronts.
Internal process agents automate the delivery pipeline:
01 DISCOVERY AGENT interviews sponsor, generates PRD 02 REQUIREMENTS ANALYST AGENT user stories, risks, preliminary architecture 03 POC BUILDER AGENT generates navigable Lovable prototype 04 QA AGENT accessibility, flow validation, GitHub issues
Every agent output passes through human review before advancing. Human-in-the-loop is not a feature. It is the contract.
Secretariat-facing agents are composable primitives for each secretariat’s own solutions:
CONSULTATION AGENT searches via File Search, answers with cited sources ANALYSIS AGENT multi-step reasoning over complex documents AUTOMATION AGENT executes workflows, fills forms, dispatches notifications
Human supervision is risk-graded by design. Low-risk informational consultation has optional review. Medium-risk drafts and analyses have recommended review. High-risk decisions affecting rights, health, or safety require mandatory human approval before any agent output reaches a citizen.
The first wave of the platform is live. Wave 0a delivered the Core (Gateway + File Search + auth + first end-to-end test) in 3–4 weeks. Wave 0b (RBAC, CI/CD, traces, ledger) ran in parallel with Wave 1 kickoff and is now complete. Wave 1 (1 to 3 production pilots) is in execution, with the first piloto targeted for production within the four-month executive window.
Beyond the technical delivery, three structural shifts.
The marginal cost of the next use case dropped substantially. New AI initiatives now compose from a shared platform instead of starting from zero. Same Gateway. Same governance. Same agent primitives. Each new use case multiplies the value of the platform instead of fragmenting it.
Governance moved from afterthought to fabric. RBAC, audit trails, and consumption ledger were designed in from week one. The platform is the first in the state’s public sector to operate with full traceability of every AI request — by project, by user, by cost.
The agent layer became durable. The four Quantum Leap process agents (Discovery, Requirements Analyst, PoC Builder, QA) became reusable across subsequent engagements. The pattern of automating consulting delivery while delivering consulting is now a core Quantum Leap primitive.