← Back to selected work

Project notes · Live

nottoobad

nottoobad is a purchase advisor that turns an open question into evidence, alternatives, and a buy, wait, or skip verdict. It uses grounded AI and keeps commercial inputs outside the decision.

Visit nottoobad
Role
Product, data, and engineering
Period
Current
Status
Live
Disciplines
Product strategy, Decision intelligence, Applied AI, Data architecture, Cloud engineering, FinOps

Product demo · 0:21 · Silent walkthrough

See a purchase decision unfold

A $1,200 espresso machine is tested against routine, evidence, alternatives, and a final Skip verdict.

nottoobad · See a purchase decision unfold

An editorial walkthrough built from first-party captures: nottoobad tests the buyer's routine, keeps the dossier inspectable, compares a simpler alternative, and reaches a structured Skip verdict.

01 · Context

What prompted it

Purchase systems usually optimize discovery and conversion. nottoobad asks the prior question: whether the purchase should happen at all.

That creates a different risk profile. Evidence is heterogeneous and untrusted, personal context evolves during the conversation, commercial incentives can conflict with judgment, and every research or model call adds latency and provider spend.

02 · Approach

How it works

Build a progressive dossier as the question develops, preserving stated context, product evidence, review themes, sources, alternatives, trade-offs, and the final verdict as one inspectable decision record.

Treat server-side product and review research as an untrusted input boundary: normalize sources, sanitize retrieved content, preserve null outcomes, and bound execution when evidence or checks fail.

Keep the verdict contract separate from commercial inputs, then protect execution with server-side user-scoped authorization and data access, idempotency, layered rate controls, provider-wallet limits, and atomic usage accounting.

03 · Decisions

Choices along the way

01

Shipped

Make the dossier the source of decision context

Constraint
A purchase question changes as needs, evidence, and alternatives emerge; a single prompt cannot preserve that evolution.
Alternatives considered
A one-shot recommendation or a fixed product score.
Decision
Accumulate stated context, normalized evidence, sources, alternatives, and verdict inputs in a progressive dossier.
Effect
The system supports revisable judgment and keeps the path to the verdict inspectable instead of collapsing it into one score.
02

Shipped

Treat research and model execution as untrusted

Constraint
Retrieved pages and generated outputs can be malformed, incomplete, adversarial, or unsupported by usable evidence.
Alternatives considered
Pass retrieved content directly to the model and accept the generated answer when execution completes.
Decision
Normalize sources, sanitize content, bound execution, preserve null outcomes, and reject the affected operation when checks or evidence fail.
Effect
Failure becomes an explicit system state, limiting how far untrusted evidence can influence the dossier or verdict.
03

Shipped

Separate judgment from incentives and spend

Constraint
Commercial value, shared cloud resources, retries, and provider costs can distort either the verdict or the operating model.
Alternatives considered
Blend commercial ranking into recommendation generation and meter usage after execution.
Decision
Design the verdict contract to keep commercial inputs outside verdict generation, with server-side user-scoped authorization and data access, idempotency, layered rate controls, provider-wallet limits, and atomic usage accounting.
Effect
The architecture bounds incentive influence, duplicate work, unauthorized access, and unaccounted provider spend at explicit control points.

04 · Current state

What is true now

Shipped
The deployed flow accepts open purchase questions and builds a dossier with server-side research, alternatives, and a structured buy, wait, or skip verdict.
Evidence
The implementation contains source normalization and sanitization, null-result handling, bounded model operations, a separate verdict contract, server-side user-scoped authorization, idempotency, rate controls, provider-wallet limits, and atomic usage accounting.
Unresolved
Recommendation quality across categories and provider behavior under sustained load remain unmeasured.
Visit nottoobad

05 · Lessons

What I learned

Trustworthy decision intelligence depends on provenance, incentive boundaries, authorization, and cost controls, not model fluency alone.

  1. Durable decision structure matters more than a fluent answer

    Evidence
    The dossier preserves the question, evolving context, normalized evidence, alternatives, and verdict inputs across the interaction.
    Next time
    Define the decision record and its revision rules before choosing prompts or models.
  2. Grounding is an operating boundary

    Evidence
    Research passes through source normalization, sanitization, null outcomes, and bounded execution before it can affect the dossier.
    Next time
    Specify evidence rejection, provenance, and degraded modes alongside the happy path.
  3. Trust and AI FinOps share the same control plane

    Evidence
    User-scoped authorization, idempotency, layered rates, provider-wallet limits, and atomic usage accounting constrain both access and model spend.
    Next time
    Design incentive, authorization, and cost boundaries together rather than adding metering after launch.