H Delivery CommandGoverned project recovery
Ahmad Bukhari + AiXCEL Solutions

Portfolio deployment · synthetic data · no external changes

Agency delivery risks,
resolved before they compound.

Six specialist agents inspect schedule, capacity, margin, creative quality, approvals, and evidence in parallel. A recovery planner proposes the next action. Deterministic policy and a human decision gate keep every operational change outside the public system.

Run a project packet See the design logic Try with your data
LangGraph state machine Role protected margin Replay mode at $0 Approval checkpoint OpenTelemetry ready

Live decision workbench

Change the evidence.
The council changes its decision.

This is connected to the deployed API. Select a preset or edit the controls, then compare the status, evidence, action, privacy boundary, and agent trace.

01

Project evidence

Editable

Replay mode uses deterministic logic and makes no model call. No staffing, invoice, assignment, publication, or client message can be executed.

02

Decision record

Ready

Awaiting project evidence

Run the council to generate a traceable recovery decision.

Architecture decisions

Reasoning where judgment helps.
Code where rules matter.

The system uses the smallest defensible boundary for each responsibility. The decision record stays inspectable, testable, and reversible.

HRMNY multi-agent collaboration diagram
01

REST and FastAPI

Typed browser friendly contracts, generated OpenAPI, and direct Postman validation. GraphQL and gRPC add no useful capability to this bounded workflow.

02

LangGraph

Parallel analysis, durable state, explicit branching, and resumable human checkpoints. A free form group chat would be harder to audit and reproduce.

03

Deterministic policy

Dates, percentages, role access, margin math, and release gates remain code. An LLM may rewrite a verified memo, but it cannot calculate or authorize.

04

No automatic mutation

The portfolio has no assignment, payroll, invoice, message, or publishing tools. Approved proposals become records until a private HRMNY integration passes UAT.

Evaluation and observability

Proof is part of the product.

Golden scenarios cover the business rules, privacy boundary, approval flow, failure behavior, and operating budgets. Traces expose behavior without capturing raw project content by default.

Current evaluation

LoadingChecking the deployed suite

Correctness

Schema, expected state, correct action, schedule variance, capacity forecast, margin math, quality gate, and approval window.

Governance

Role protected margin, evidence grounding, approval enforcement, unauthorized action blocking, and prompt injection containment.

Operations

Idempotency, latency, step budget, zero cost replay, consistency, error type, retry count, and state transitions.