I build AI systems for real business workflows — combining agentic reasoning with deterministic software, enterprise integrations, evaluation, reliability, and human oversight.

My interest is at the intersection of:

Applied AI × Systems Engineering × Business Operations

   

I am particularly interested in AI systems that have to operate inside existing business processes, rather than isolated chat interfaces.

That means dealing with:

• messy and heterogeneous data
• APIs and enterprise systems
• legacy infrastructure
• ambiguous operational cases
• business rules and policy
• human approval
• failure and recovery
• security and authorization
• evaluation and regression
• measurable operational outcomes

My current work explores this through three different problem classes.

AI handles ambiguity. Deterministic software handles truth, authority, and execution.

LLMs are powerful reasoning components, but they should not automatically become the system of record, authorization layer, or source of truth.

I therefore design around explicit boundaries:

The objective is not to make an AI system look autonomous.

It is to make it useful, bounded, observable, and reliable enough to participate in real operational workflows.

View Repository →

FinSight is a focused financial-resolution system modeled around a single B2B commerce company.

Its job is to:

Detect → Investigate → Explain → Propose → Authorize → Execute → Verify → Close

The system reasons across heterogeneous financial systems including:

• payment-provider state
• accounting records
• expected settlement data
• email and collaboration context
• a legacy COBOL-style settlement boundary

                Financial Systems
                       │
        ┌──────────────┼──────────────┐
        ▼              ▼              ▼
    Payments       Accounting      Legacy
        │              │              │
        └──────────────┼──────────────┘
                       ▼
              Deterministic Core
          Reconciliation · Evidence
             Rules · State · Policy
                       │
                       ▼
                AI Investigator
             Reason · Correlate
                 · Explain
                 · Propose
                       │
                       ▼
                Human / Policy
                    Gate
                       │
                       ▼
                   Execute
                       │
                       ▼
              Independent Verify
                       │
                       ▼
                     Close

The agent can:

• investigate
• gather evidence
• generate hypotheses
• correlate information
• explain discrepancies
• propose resolutions

The agent cannot:

• establish financial truth
• authorize its own action
• bypass policy
• execute arbitrary operations
• verify its own execution

Current engineering work includes:

• deterministic financial reconciliation
• evidence-backed investigation
• typed agent capabilities
• authority boundaries
• human resolution workflows
• legacy batch integration
• idempotent execution
• adversarial evaluation
• security and isolation
• post-execution verification

Status: Active development

View Repository →

ClaimOps explores how AI can assist claims operations while keeping adjudication authority with the insurer or TPA.

The core workflow:

Claim
  ↓
Document Ingestion
  ↓
Classification / Extraction
  ↓
Evidence & Provenance
  ↓
Deterministic Validation
  ↓
Exception
  ↓
Bounded Investigation
  ↓
Evidence-Grounded Finding
  ↓
Human Review
  ↓
Audit

• interpreting heterogeneous documents
• investigating ambiguous exceptions
• gathering relevant evidence
• forming hypotheses
• preparing findings

• validation
• state transitions
• authorization
• evidence verification
• workflow execution
• auditability

The system is deliberately positioned beside the insurer's adjudication process rather than replacing it.

Current engineering work includes:

• evidence-grounded agent tools
• deterministic-first orchestration
• durable workflow state
• failure and retry semantics
• tenant isolation
• adversarial testing
• evaluation harnesses
• observability
• human-review routing

Status: Active development

View Repository →

OntologyAI explores the problem that comes before implementation:

How do you turn messy business context into a structured understanding of an organization, its processes, systems, and operational problems?

The core idea:

Business Context
       ↓
Domain Model
       ↓
Entities & Relationships
       ↓
Processes
       ↓
Operational Pain Points
       ↓
Solution Design

It focuses on the discovery and solution-design side of enterprise AI engineering.

Status: Experimental / evolving

These projects are intentionally different.

They explore different business problems, data shapes, system constraints, and risk models.

But they share the same engineering principle:

Business Problem
       ↓
Workflow Understanding
       ↓
System & Data Mapping
       ↓
Deterministic Controls
       ↓
Bounded AI
       ↓
Human / Policy Boundary
       ↓
Execution
       ↓
Verification
       ↓
Observable Outcome

I am interested in where AI belongs inside a system — and equally, where it should not be trusted.

Python · FastAPI · LangGraph · LLM APIs · RAG · Tool Calling · Structured Outputs · Context Engineering · Agent Evaluation

PostgreSQL · Redis · Pydantic · AsyncIO · REST APIs · Webhooks · SQL · Schema Mapping · Data Reconciliation

APIs · Events · Queues · Batch Processing · Object Storage · Legacy Systems · Canonical Models · Idempotency

TDD · CI/CD · Observability · OpenTelemetry · Auditability · RBAC · Isolation · Prompt-Injection Defense · Failure Testing

AWS · GCP · Docker · Linux

I prefer an evidence-driven development loop:

Specification
     ↓
Contract
     ↓
RED Tests
     ↓
Implementation
     ↓
GREEN
     ↓
Integration Testing
     ↓
Adversarial Testing
     ↓
Review
     ↓
PR
     ↓
Merge
     ↓
Evidence

A system is not complete because the happy path works.

I want to understand:

• What happens when data is malformed?
• What happens when the model is wrong?
• What happens when a tool fails?
• What happens when a request is duplicated?
• What happens when evidence is missing?
• What happens when state becomes stale?
• What happens when an external dependency disappears?
• What prevents an agent from exceeding its authority?

I am developing toward roles at the intersection of:

Applied AI · Agent Engineering · AI Solutions · Enterprise Integration · Forward-Deployed Engineering

I enjoy problems where the work starts with an ambiguous business process and ends with a working technical system:

Understand the business
        ↓
Map the workflow
        ↓
Understand the data
        ↓
Identify the real constraint
        ↓
Design the solution
        ↓
Build the system
        ↓
Integrate with existing infrastructure
        ↓
Evaluate it
        ↓
Deploy and observe it
        ↓
Measure the outcome

I am particularly interested in the question:

How do we make AI useful inside real organizations without pretending that probabilistic models are deterministic systems?

That's the engineering problem I'm exploring.