Miral Shah
Co-founder
Miral Shah is a co-founder of Humint Labs, a Melbourne-based enterprise AI company, and leads its AI, technology, strategy and operations. Miral heads the Applied AI team that designs, builds and runs production AI for contact centres and enterprise service. Before Humint Labs, Miral led enterprise customer success for contact centre and digital customer experience platforms. Miral writes about where AI earns its place in customer service, and where people should stay in charge.
Articles by Miral
Part 5 of 5 in the series Where a decision model fits in enterprise service workflows
How we tested decision models
The harness, governed scenarios, answer key and scoring controls we built to compare decision models repeatably.
Part 4 of 5 in the series Where a decision model fits in enterprise service workflows
Putting decision models to work in the contact centre
Four places a decision model earns its keep in the contact centre and digital service, what a person still decides, and what each design costs to run.
Part 3 of 5 in the series Where a decision model fits in enterprise service workflows
How decision models compare in the contact centre
Which decision models are good enough for contact centre routing and handoff, what they cost, and how to choose between them, from a repeated Humint Labs test.
Part 2 of 5 in the series Where a decision model fits in enterprise service workflows
How we tested the Jev decision model, step by step
The cases, scoring rules and limits behind our Jev decision model test, written so another team can design a comparable test.
Start of a series, part 1 of 5
Where a decision model fits in enterprise service workflows
Where a fast decision model fits in your service, where it does not, and how to trial it safely, from a repeated Humint Labs test.
Part 3 of 3 in the series Where Copilot stops and a custom agent starts: drawing the line in a Microsoft estate
The operating model for a custom AI agent
The ownership, evaluation, incident and lifecycle responsibilities that turn a custom AI agent into an accountable service.
Part 2 of 3 in the series Where Copilot stops and a custom agent starts: drawing the line in a Microsoft estate
Copilot data governance readiness in a Microsoft estate
The information architecture, permissions and content-management checks that should precede a Copilot rollout.
Start of a series, part 1 of 3
Where Copilot stops and a custom agent starts: drawing the line in a Microsoft estate
A decision framework for choosing a licensed Copilot, configured agent or custom agent based on data, authority and operating responsibility.
Part 3 of 3 in the series Prompt injection when the model has tools
Testing prompt-injection defences with release evidence
How to test prompt-injection controls without over-claiming detector performance or treating one exercise as a permanent result.
Part 2 of 3 in the series Prompt injection when the model has tools
Prompt injection provenance and capability controls
How to classify untrusted inputs, verify authority outside the model and constrain the capability union available to a tool-using agent.
Tracing a conversation end to end: model calls, guardrail decisions and routing
What a span should hold in a conversational system, why the session is the unit of tracing, and what it costs to watch. From our own instrumented front door.
Start of a series, part 1 of 3
Prompt injection when the model has tools
Why filtering is not enough once a model can access tools, data or consequential actions, and how to structure the control problem.
Part 4 of 4 in the series Choosing models like a portfolio, not a religion
AI model lifecycle and deprecation: plan the exit before it is urgent
A governance approach for model retirement, open weights, lock-in boundaries and accountable ownership.
Orchestrating a multi-agent claims workflow: topology, handoffs and blast radius
Four agents given a shared goal will find an answer. Ask them six months later how they found it and you get a transcript, not a decision record.
Part 3 of 4 in the series Choosing models like a portfolio, not a religion
Testing AI model substitution before a provider change forces it
Evidence and release gates for replacing an AI model without mistaking a benchmark result for workload readiness.
Part 2 of 4 in the series Choosing models like a portfolio, not a religion
AI model routing and fallback: separate decisions, shared evidence
How to route enterprise AI workloads between models, design fallbacks and measure both paths without conflating their purpose.
Start of a series, part 1 of 4
Choosing models like a portfolio, not a religion
A practical way to select and govern a small portfolio of models, with clear routing, substitution and exit decisions.
One governed toolset, three assistants: MCP servers as enterprise infrastructure
Four assistants, one management surface, and the layer that decides what any of them may actually do.
Running an evaluation harness as production infrastructure on AWS
Evaluation suites decay rather than fail. Versioned corpora, a run manifest, separated scoring stages, a queryable history and a release gate on AWS.
Inside the agent evaluation arena: how we score autonomous systems
We built a simulated contact centre to pit multi-agent language systems against reproducible tasks. Here is the harness, the metrics that actually predict production reliability, and where today's models keep tripping.
Building a retrieval pipeline that holds up in production
Chunking, embedding and retrieval on Postgres and pgvector that stays fast and accurate once real documents and real users arrive.
Testing language-model features before they meet users
A practical evaluation method for language-model features: representative scenarios, calibrated graders, adversarial testing and release evidence that survives change.
A practical framework for enterprise AI governance
Governance is a design problem before it is a compliance one. Here is how we build guardrails that engineering teams actually use, grounded in ISO 42001 and the direction Australian regulation is now taking.
Using simulation tooling to train and assure AI agents
How Humint Labs has used custom simulation apps across delivery engagements to train and assure chat, messaging and voice agents, from early feasibility work through to repeatable release evidence.