ABOUT MUDANZA

Memory that serves the relationship.

Mudanza is an early-stage advisory platform and governed memory layer for wealth management. We are building around a simple idea: what a firm learns about its clients should carry forward, with the context and evidence that make it useful.

A client’s priorities live across conversations, notes, documents, and separate systems. Our direction brings that knowledge into one living record, connecting it to planning and the work that follows. AI should help people prepare, understand, and follow through while preserving human judgment.

Our five co-founders bring experience in advisory practices, capital markets, technology consulting, enterprise data and AI, and building businesses. We are still early: these roles will help shape both the product and the way we build it.

Why this role matters

The value of a living client record is what a team can safely do with it. A meeting note, a family update, and a planning decision should inform the next conversation without losing their source or becoming unquestioned facts. You will build the retrieval, memory, and reasoning workflows behind that experience. The challenge is to make assistance useful while preserving uncertainty, permission boundaries, and the distinction between a proposed action and an executed one.

The scope of your role

This is a production engineering role, working closely with the co-founders and the product and data engineers. It includes model behavior, application integration, evaluation infrastructure, and operational ownership. You will help decide when a model is useful, when a deterministic read is more appropriate, and when the system should ask a question or explain that evidence is unavailable. Quality, latency, privacy, and cost are all part of the work.

What you’ll own

Build retrieval and memory pipelines that connect meetings, notes, documents, and structured records while preserving who supplied information, when it was learned, and where it came from.
Handle changing facts, conflicting sources, corrections, and stale context. Design updates so an earlier interpretation does not silently overwrite stronger evidence.
Develop firm-, advisor-, and client-scoped assistance, with server-enforced tool permissions and explicit review for consequential actions.
Create representative evaluation sets for answer grounding, task completion, scope isolation, refusal and recovery behavior, and malicious instructions embedded in retrieved content.
Improve task-aware model routing within approved data-handling constraints. Measure quality and latency before choosing a cheaper or more capable route.
Instrument usage and cost by feature and request context. Investigate outliers, long conversations, repeated tool calls, and failures that waste both time and tokens.
Build observable workflows with bounded retries, traceable tool results, and clear stop conditions. Work with product design to make sources, limitations, and review status understandable.

What you bring

Experience shipping LLM-powered features beyond a prototype, including evaluation, monitoring, and investigation of real failures.
Strong software engineering skills in TypeScript or Python and the ability to integrate model workflows into an existing application.
Practical understanding of retrieval, tool calling, structured outputs, context management, and the ways generated answers can be misleading.
The ability to construct useful evaluation cases, compare against a baseline, and separate anecdotal improvement from repeatable evidence.
Experience handling sensitive data and designing systems where authorization and human approval cannot depend on the model following an instruction.

Especially useful experience

These are helpful additions, not a checklist you must complete.

Azure AI services, multi-model routing, retrieval over financial documents, or production voice and transcription workflows.
Experience with adversarial evaluation, long-lived memory, audit records, or usage-based cost attribution.

Your first 90 days

A starting framework for the role; we’ll shape the sequence around product and customer priorities.

Days 1–30
Map one existing AI workflow and establish a representative baseline for correctness, latency, and cost.
Days 31–60
Ship a measurable improvement with regression coverage for cross-client context, unsupported claims, and action boundaries.
Days 61–90
Put evaluation and operational feedback into the release process, with clear criteria for rollout, rollback, and further iteration.

How we approach the work

Start with the person and the problem. Make a small, complete improvement. Keep sources and assumptions visible. Test the important failure cases, and follow the work through to its actual use. We value clear writing, thoughtful disagreement, and the ability to change direction when the evidence changes.

In an early team, ownership includes the edges: documenting a decision, helping a teammate, investigating an issue, and being honest about what is ready and what is still being built.

How to apply

Walk us through an AI feature you operated. Include a failure you discovered, the evaluation that exposed it, and how you balanced answer quality with latency or cost. Use sanitized examples only.

You do not need a particular degree, employer pedigree, or a perfect match to every useful skill. Explain the relevant work you have done and the perspective you would bring. If an adjustment would help you participate in a hiring conversation, mention it in your introduction.

Applied AI Engineer