Systems Architecture Library

Engineering concepts explored in my projects, prototypes, and research.

Browser Workflow Discovery & Replay

Using an LLM to discover browser workflows once, then saving successful workflows for deterministic, model-free replay.

Key Design Concepts

LLM-guided discoverySaved workflowsDeterministic replayModel-free execution

Demonstrated In:

Capability Runner

Separates workflow discovery from replay so saved browser workflows can run without repeated model calls.

Change Data Capture & Data Synchronization

Propagating database changes through a distributed pipeline for downstream caching and indexed search.

Key Design Concepts

PostgreSQL change captureDebezium connectorsKafka data propagationRedis and Elasticsearch

Demonstrated In:

SyncStream

Connects PostgreSQL, Debezium, Kafka, Redis, and Elasticsearch in a change-data-capture pipeline.

Voice Interactions & Agent Orchestration

Combining speech recognition and WebSocket-based voice interactions with retrieval and agent orchestration for an IT/HR support prototype.

Key Design Concepts

Speech recognitionWebSocket interactionsLangGraph orchestrationEnterprise RAG

Demonstrated In:

Ticketless IT/HR Voice Support

Brings voice interactions, enterprise knowledge retrieval, and support workflow orchestration into one prototype.

Action Validation & Human Takeover

Checking browser actions against page evidence and handing unsupported states back to a person.

Key Design Concepts

Action validationPage-evidence checksExpected-error handlingHuman takeover

Demonstrated In:

Capability Runner

Validates actions and page evidence, handles expected errors, and supports human takeover when automation cannot proceed.

Case-Aware Retrieval & Reusable Memory

Combining enterprise RAG with reusable case memory to support case-aware agent workflows.

Key Design Concepts

Enterprise retrievalReusable case memoryDual-memory agentsStaged evaluation

Demonstrated In:

Adaptive Runbook Intelligence Platform

Explores support-agent workflows that combine retrieved knowledge with reusable case context.

Retrieval Reranking & Context Selection

Exploring how dynamic and fixed-K reranking select retrieved context for language models.

Key Design Concepts

Retrieval-augmented generationDynamic K selectionFixed-K rerankingLLM context selection

Demonstrated In:

Dynamic Retriever

Compares dynamic and fixed-K reranking approaches for choosing context in RAG workflows.

Compute-Efficient Learning

Studying adaptive computation through conceptual criticality and measuring the trade-off between model accuracy and energy use.

Key Design Concepts

Adaptive computationConceptual criticalityEnergy measurementAccuracy evaluation

Demonstrated In:

Adaptive Compute-Efficient Learning

AAAI 2026 Student Abstract proof of concept retained about 90.7% accuracy while reducing energy use by about 65% versus a 6-layer baseline.

Screen Understanding & Support Verification

Using screen context to support IT/HR assistance, with human approval and structured verification before closing a support workflow.

Key Design Concepts

VLM screen understandingOCRHuman approvalStructured verification

Demonstrated In:

Ticketless IT/HR Voice Support

Combines VLM/OCR screen understanding with voice support, enterprise retrieval, and approval before closure.