Inspiration
Estate-sale listings can contain hundreds of poorly labeled or completely unlabeled photographs. Finding something valuable means manually going through those photos, figuring out what each object might be, reading signatures and labels, and then researching items one by one.
I wanted to know whether an AI agent could take over that entire investigation from just one estate-sale URL.
But while building Estate Lens, the problem became more interesting than simply finding valuable objects. An AI system can find evidence very easily. The harder problem is deciding whether that evidence actually proves what the agent thinks it proves.
That became the central idea behind Estate Lens:
Build an agent that investigates the evidence, challenges its own findings, and knows when it doesn't have enough proof to make a decision.
What Estate Lens Does
Estate Lens is an autonomous multimodal research agent for estate sales.
Give it one estate-sale URL and it:
- Ingests the listing and analyzes the entire photo gallery.
- Uses Gemini vision to identify objects and visual clues worth investigating.
- Separates direct observations from interpretations and keeps possible maker or artist identities labeled as HYPOTHESIS.
- Researches promising candidates using Google Search grounding and first-party auction sources.
- Fetches auction pages and verifies sold results at the source.
- Separately determines whether a verified sale is actually relevant enough to use.
- Calculates conditional value ranges only when the evidence supports them.
- Routes uncertain candidates to an independent Skeptic Agent.
- Uses deterministic rules to apply the Skeptic's findings and determine the final action.
- Publishes a ranked report showing the evidence, uncertainty, and recommended next human step.
The important distinction is that **finding a real sale does not automatically make it a valid comparison.
A Real 375-Photo Run
Our Golden Demo processed a real 375-photograph estate sale through the production Google Cloud workflow.
The run escalated 50 candidates and ultimately found 29 source-verified sold transactions.
But Estate Lens refused to use most of them for valuation.
Only 6 were admitted as relevant comparisons. Another 23 genuine, source-verified sales were retained as market evidence but explicitly excluded from valuation because they were not sufficiently relevant.
That behavior became one of the clearest demonstrations of what I wanted this agent to do: refuse its own evidence when the evidence doesn't actually support the conclusion.
One example was a possible Louis Vuitton case identified from a label visible inside the object. Estate Lens found 8 source-verified Louis Vuitton sales, but rejected 6 because they weren't good enough matches. Only 2 remained relevant.
Even then, the system wasn't satisfied.
The independent Skeptic Agent recognized that the estate-sale listing only provided an interior photograph. Important evidence such as the exterior, hardware, dimensions, and additional authentication markers was missing.
Its verdict was:
INSUFFICIENT EVIDENCE
Deterministic policy then lowered the candidate's confidence and opportunity score. The final recommendation remained:
AUTHENTICATION REQUIRED
Because the acquisition price was unknown, Estate Lens also refused to calculate ROI.
The system found evidence, rejected most of its own evidence, challenged what remained, and refused to make a stronger claim than the proof allowed.
How I Built It
Estate Lens runs as an asynchronous workflow on Google Cloud.
The production architecture uses:
Vertex AI for Gemini model access Gemini 3.5 Flash-Lite for large-scale multimodal vision triage and grounded discovery Gemini 3.7 Flash for the independent Skeptic Google Agent Development Kit (ADK) for agent orchestration Google Search grounding for evidence discovery Cloud Run for the public web application and private worker Cloud Tasks for asynchronous, authenticated execution Firestore for durable run state, stages, and timeline events Cloud Storage for durable run artifacts Cloud Build + Artifact Registry for deployment Cloud Logging / Monitoring for operations
The public Cloud Run service and execution worker are deliberately separated. The public service can display stored runs but has no Vertex AI permission. New work is dispatched through Cloud Tasks to a private worker using authenticated OIDC.
This means the expensive execution path isn't directly exposed to the public internet.
Models Propose. Deterministic Code Decides.
One of the most important architectural decisions was not allowing the language models to control the final recommendation.
Gemini can observe an image, propose an identity, generate research queries, and challenge a hypothesis.
But deterministic, versioned code controls:
candidate escalation research routing evidence eligibility valuation opportunity scoring final actions the effect of the Skeptic review
A model therefore cannot simply decide that an item is valuable or authenticated.
Possible attributions remain hypotheses, and a completed run can be replayed from its stored evidence without making new model or provider calls.
Challenges
Finding evidence was easier than trusting it
Google Search grounding is extremely useful for discovery, but search results can contain claims and prices that haven't been independently verified.
Instead of trying to solve this only through prompting, I changed the architecture.
Grounded search can discover evidence, but it cannot establish a sale price. Estate Lens must fetch the underlying source and verify what the page actually proves.
A verified sale isn't necessarily a comparable sale
This became one of the biggest lessons of the project.
A transaction can be completely real, correctly parsed, and from the correct maker—and still be terrible evidence for valuing the object in front of you.
Estate Lens therefore separates source verification from comparability.
A verified transaction can remain in the Evidence Ledger while being marked NOT A COMPARABLE and prevented from affecting valuation.
AI uncertainty doesn't always look uncertain
During development, I found that forcing a model to stop returning an uncertain classification did not necessarily improve accuracy. Sometimes the uncertainty simply moved into a more confident-looking wrong answer.
That changed how I approached model outputs: uncertainty needed to be preserved and measured rather than prompted away.
Production agent workflows need durable state
A 375-photo investigation cannot depend on the browser remaining open or a single HTTP request surviving.
Cloud Tasks, Firestore, Cloud Storage, idempotent processing, and checkpointed execution allowed the workflow to continue asynchronously and preserve its evidence across multiple worker invocations.
Accomplishments
The Golden Demo completed a 375-photo estate sale end to end on Google Cloud without manually selecting which objects the agent should investigate.
It produced:
375 photographs discovered 50 escalated candidates 29 source-verified sold transactions 6 relevant valuation comparisons 23 verified-but-excluded transactions 3 conditional value ranges 47 explicit evidence insufficiencies 3 independent Skeptic reviews, all of which lowered confidence 3,904 durable workflow events
Most importantly, the system was willing to say INSUFFICIENT EVIDENCE instead of manufacturing certainty.
What I Learned
The hardest problem in agentic research isn't finding information.
It's knowing which information should be trusted, which evidence actually applies, and when the agent should refuse to make a decision.
I also learned that deterministic systems become especially valuable around probabilistic models. Gemini is excellent at observing, interpreting, researching, and challenging. Deterministic gates give those capabilities boundaries that can be tested, reproduced, and audited.
Estate Lens isn't an authentication service or a certified appraisal. It is an investigation and opportunity-detection agent designed to tell a human not only what might be worth looking at, but why—and what still needs to be proven.
Today, we're applying that idea to estate sales.
But that's just the beginning.
Log in or sign up for Devpost to join the conversation.