01 / Contracted tools
Tools declare what they establish.
Versioned contracts define parameters, coverage, sensitivity, freshness, caveats, and payload shape. The agent receives an a-la-carte menu rather than ambient database access.
Case study 02 · Enterprise architecture & AI
Enterprise Intelligence Platform
Shared enterprise data, governed tools, and AI agents form a common foundation for research, decision support, and operational workflows. Developed from a production federal opportunity system, the platform architecture keeps evidence, authorization, and operational ownership inside the runtime.
- Role
- Product architect, agent-system designer, delivery lead
- Domain
- Federal opportunity and enterprise decision intelligence
- Runtime
- AWS predecessor; enterprise Azure and Databricks successor
- Boundary
- Described without customer or employer detail; the live predecessor and the enterprise architecture are distinct evidence states
The problem
Decisions were scattered across systems and people.
Opportunity discovery was only the first step. A useful system had to connect solicitations, amendments, awards, competitors, employee capabilities, proposal evidence, and the judgment used to decide what deserved pursuit.
The source data changed at different cadences and carried different authority. The workflows lived partly in databases, partly in documents, and partly in experienced operators. A model with a large prompt could produce fluent answers, but fluency did not make those answers current, authorized, reproducible, or safe to act on.
The work therefore expanded from an application into a platform problem: create a durable data and tool boundary that many agent products could share, while keeping identity, row scope, sensitivity, freshness, and claim limits visible at run time.
What I built
From opportunity app to governed agent platform.
The public predecessor proves the workflow. The enterprise architecture separates durable data, contracted tooling, agent execution, and disposable product surfaces.
02 / Corrective retrieval
Retrieval is an observed loop.
Search quality is measured and corrected instead of hidden behind one semantic query. Structured evidence is separated from synthesized narrative and unsupported claims are surfaced.
03 / Runtime governance
Authorization travels with the run.
Identity, capability grants, sensitivity, and row scope shape every dispatch. Results carry evidence labels upward; audit and cost attribution make the decision path inspectable.
Architecture
A durable base for disposable applications.
Applications depend on stable runtime, tool-menu, and renderer contracts. The platform remains while products change.
- Sources — enterprise and public data
Opportunity, award, workforce, CRM, funding, document, and user context.
- Governed data
Source-mirrored and platform-owned state, classification, lineage, and refresh ledgers.
- Compiled tool menu
The typed read path with contracts, guarded dispatch, and evidence envelopes.
- Agent runtime rail
Identity, authorization, model abstraction, evaluation, telemetry, and audit.
- Decision surfaces
Chat, briefing, campaign, dashboard, profile, and explorer experiences.
The evidence envelope — source reference, lineage, and approval state — is the object that travels between every stage. It is what lets a decision surface show where an answer came from and whether it was approved.
Design principle
The agent should never know more authority than the tool result can prove.
What this demonstrates
- Workflow discovery and product decomposition
- Data modeling with lineage and freshness
- Agent tool contracts and runtime design
- Authorization, row scope, and auditability
- Provider abstraction and streaming UX
- Container, cloud, and operating-team handoff
What it cost
The option I rejected, and what the choice took.
Senior leaders were split on what the enterprise platform should be first. The decision was not to choose between them.
Described without customer or employer detail.
Rejected
Pick a side.
Build the quick tools the operators asked for first, or the strategic view the executives asked for first. Each camp had a budget argument and a sponsor.
Chosen
The shared data first.
One fabric both products would read from, then a first cut at the strategic view while the operator tools were built in parallel on it. Each makes the other better.
Cost
Months before either camp saw its product.
The fabric had to earn its keep with prototypes and early tool wins while it landed. Sequencing is only credible if the wait is visible and short.
Outcome
Working proof, then a platform boundary.
The result is two complementary forms of proof: a live predecessor that turns federal opportunity data into an operating workflow, and an enterprise architecture that makes new agent products installable on shared governed rails.
The most important outcome is not a permanent tool count. It is the repeatable separation of source authority, application transformation, agent capability, user authorization, and operational evidence, so teams can change one without silently changing the others.