NAVIDBR

  1. loading sources
  2. verifying records
  3. checking boundaries
  4. 04/04 PASS

NAVIDBRApplied AI Systems

Navid Broumandfar

making systems that actually work

naming systems that quietly break

Applied AI systems. Prototypes and labs, each status written down.

Solo systems. Prototypes and labs, each gap written down.

About me

I’m an applied AI builder

turning repeated work into governed systems. Sources, boundaries and proof stay visible.

Building in public, solo. Records link to runnable code.

What i do

  • Workflow

    I start from the repeated work: who owns it, where it hands off, and where it quietly breaks.

  • Evals

    Approval, refusal, escalation and regression cases, written down before an agent gets more scope.

  • Proof

    Public records with status labels and source limits, for the parts demos usually skip.

Records

Public proof

for the parts that demos usually skip: sources, limits, evaluation and failure paths.

for prototypes and labs, built solo, with sources, limits and boundaries written down.

  • PROOF CaseOps Lakehouse Databricks evidence preparation the unglamorous data prep, documented
  • PROOF CaseOps Control Tower AWS Bedrock grounded review grounded review, escalations included
  • LAB Agent Behavior Evals Lab agent behavior evals 98 percent of my cases caught, 21.8 blind GitHub
  • LAB Agent Terrarium sealed, observable agent operations autonomy inside walls it cannot cross
  • PROOF Agent Gate Demo a CI gate for unsupported agent claims the pipeline refuses before I do
  • PROTO Evidence Loop evidence-first visibility proposals four releases, a thin test surface
  • PROTO Purchase-Intention MLOps e-commerce decision thresholds thresholds chosen before the demo

Principles

  • Workflow

    Before the model, there is the workflow: the repeated work, its owner, the handoff, the failure path.

    If I can’t name the owner, the handoff and the failure path, the model choice doesn’t matter yet.

  • Boundaries

    Autonomy earns scope only when sources, allowed actions, refusals and escalation are visible.

    An agent that can’t show its sources, refuse or escalate isn’t autonomous. It’s unsupervised.

  • Proof

    A claim gets stronger when the artifact, its status label and the remaining gap are all visible.

    A demo that hides its gaps is marketing. Show the artifact, its status label and the gap.

  • Evals

    Approval, refusal, escalation and regression cases are written before an agent gets more scope.

    The eval suite is the job description. If it isn’t written down, the agent doesn’t get the scope.

  • Moat

    If the model vendor ships your feature free in twelve months, what is still worth paying for?

    Model capability is rented. The workflow, the data and the trust are the parts you actually own.

Loop

WHAT THE LETTERS SAY

  • Before the model, there is the workflow

    Behind the demo, there is the cleanup

    Public letter2026-06-21navidbr.me/letters

  • Boundaries before autonomy

    Unbounded agents overreach

    Public letter2026-06-20navidbr.me/letters

  • Proof before product claims

    Proof arrives after claims

    Public letter2026-06-19navidbr.me/letters

Products

The products I ship and the worlds I build, kept open, each with its status written down. More are in construction.

WHAT EVERYONE SAYS

Ideas are easy

Heard everywhere

MY MOTTO

Proof is hard

Navid Broumandfar

Connect