Buddy Guzmán
I investigate how software architecture behaves under real operational constraints. What excites me most is turning architectural questions into systems I can build, test, compare and refine — then sharing what the evidence teaches me.
Current lines of inquiry
Architecture, distributed behavior, operational transformation and reproducibility.
Strong module boundaries without paying the distributed-systems premium too early.
The reference implementation uses a single deployable modular monolith with hexagonal boundaries. The technical question is not monolith versus hexagonal architecture; those describe different concerns. The interesting boundary is where deterministic decision authority ends and replaceable infrastructure begins.
Technical note
Determinism here is structural, not merely conventional: canonical ordering, isolated verifier dependencies, injected time, stable hashing, immutable snapshots and replay verification remove nondeterministic inputs from the authoritative path.
- The verifier remains isolated from agents, JPA, network and SQL APIs.
- JPA is constrained to adapters.
- Ports exist at seams where replacement is real, not as ceremony between every internal class.
Redesign the operating model before digitizing the existing friction.
This thesis line studies public-education operations around UGEL 04: institutional responsibility, process handoffs, controls, decision rights and the difference between automation and actual operating-model transformation.
Technical note
The research model separates institutional governance from operational execution. SUNEDU is therefore treated as an adjacent regulatory reference under scope validation rather than being forced into the MINEDU → DRELM → UGEL operational chain.
Uncertainty should be modeled as state, not hidden behind a confident output.
This line studies evolving computational hypotheses: how signals become observations, how confidence changes, how revisions are preserved, and how an experiment can expose its own uncertainty without pretending to be authoritative.
Exactly-once transport does not guarantee exactly-once business effect.
The hard problem starts when a reliable event crosses a transaction boundary and touches an external side effect. Idempotency belongs where the business invariant can actually be protected.
External data changes faster than the domain model should.
This queued case study examines a Node.js/TypeScript backend where external data, validation, pagination, pricing checks and reporting cross several service boundaries. The public write-up will remain architecture-only until source material is fully sanitized.
Reproducibility is an architectural property, not a deployment afterthought.
A clean-room Python example is planned around event orchestration, retries, idempotency, observability and infrastructure as code using synthetic data and independently reproducible behavior.
When knowledge becomes executable state, semantics need versioning too.
Agentic systems increasingly act against enterprise knowledge, policy and tool permissions that evolve independently. This research thread studies how to preserve decision authority when the semantic context itself changes over time.
Technical note
Replayable agent decisions require more than storing the prompt and output. A meaningful replay may need the input, model revision, ontology revision, policy revision and tool-authorization state that were active when the action was evaluated.
- Semantic drift becomes an architectural concern once knowledge can authorize behavior.
- Policy evaluation should be pinned to an explicit revision rather than “whatever is current”.
- Provenance and temporal validity make historical decisions explainable without reinterpreting them through today’s semantics.
Production reasoning
When tests pass and reality does not
Before touching the implementation, reconstruct the smallest failing path and challenge the invariants behind the expected behavior. Code expresses intent; production evidence reveals where that intent stopped matching reality.