LastTube

Inspiration

A discontinued lipstick is not just a missing SKU. The replacement advice scattered across videos, forums, and retailer pages goes stale, while a shade that looks close in a product photo can look different on a face. We wanted one narrow workflow that answers the decision a shopper actually has: what was actually listed, which images contain usable shade evidence, and does any candidate survive a same-face decision with enough evidence to act?

Public user reports describe searches lasting years and several poor matches, and stale availability references. These anecdotes support the problem, not LastTube's accuracy, product validation, market size, or willingness to pay.

What it does

LastTube starts with a remembered discontinued shade. SerpApi turns a Google Shopping search into structured candidate evidence—exact listing title, merchant, price, observed offer/source URL, and observation time. The shopper shortlists up to three candidates. Every shortlisted item must finish as either a recorded system exclusion or an explicit human accept/reject. LastTube rejects merchant images below 10% usable saturated foreground coverage, requires a successful Perfect Corp render of the remembered shade as the comparison baseline, and renders passing estimates on the same sample face. Preference is accepted-only. CIE76 explains the difference; it never chooses or restores a rejected item.

The action gate models listing identity, exact variant, shade, finish, source-image hash/coverage, and candidate VTO proof as separate present, absent, or unknown evidence. An observed offer becomes actionable only when all required fields, the baseline, and human preference are present.

In the recorded path, Ngozi Mauve Rose Matte Lipstick is system-excluded at 2.5% usable coverage; the human rejects lower-ΔE Anastasia Beverly Hills Lip Velvet, accepts and prefers NYX Professional Makeup Fat Matte Lipstick, and therefore visibly overrides the metric. The preferred has no retained trusted shade or variant record, its source-image bytes/hash were not retained, and its candidate VTO has task/poll metadata rather than a lifecycle receipt. LastTube displays those gaps, ends with No actionable lead yet, offers a refined-search handoff, and exposes no Buy action or formulation claim.

Every provider result is visibly stamped LIVE, FIXTURE, UNAVAILABLE, or FAILED. Listings are described as observed evidence, not real-time stock guarantees. Candidate colors and CIE76 ΔE are explanation aids over approximated shades, not promises of real-world appearance. When evidence is missing, LastTube stops instead of inventing a verdict.

How we built it

• React 19 and Vite render the three-act beauty-editorial interface: the loss → the hunt → the verdict.
• A Hono/Node service keeps sponsor credentials server-side and exposes typed API boundaries.
• SerpApi's google_shopping engine supplies timestamped structured listing evidence.
• Perfect Corp's Makeup VTO follows the real asynchronous lifecycle: create a task, poll inside a bounded budget, download the signed result before expiry, and retain its exact output bytes.
• Sharp estimates a dominant saturated product color behind HTTPS, hostname, response-size, and 10%-minimum usable-coverage guards. Current live estimates return their exact fetched input hash and byte count. Low-coverage images fail closed. Shared color utilities convert approximated hex values to Lab D65 and compute CIE76 ΔE.
• A deterministic demo mode replays a receipted SerpApi response, one receipted lost-shade baseline, and candidate outputs with explicitly narrower metadata-only provenance. It is labeled FIXTURE on every affected surface and guarded against accidental live calls.
• New live searches open a versioned evidence run from the SerpApi wire digest and a separately hashed, sanitized raw response retained before normalization. Shade estimation retains the exact merchant-image bytes. Candidate VTO retains every sanitized create/poll response with timestamps, safe URL lineage, request inputs, and downloaded output bytes. Validation re-derives the listing and re-hashes four artifacts before review. The current host is ephemeral, so this is per-run export—not durable storage and not proof of any current fixture candidate.
• A structured enrichment contract defaults exact variant/shade/finish to unknown and has no title parser. Only a future receipt-bound trusted record can promote those fields; no live enrichment source or actionable lead is claimed.
• Pure shared policy derives candidate dispositions and the action boundary. A clearly labeled synthetic policy fixture proves the theoretical all-fields-present branch; it is not genuine product evidence. Real incomplete and unknown branches fail closed.
• The production adapter serves the built client and API from one container; public deployment is intentionally an operator gate.

Where the sponsor technology does the real work

Perfect Corp: Makeup VTO is the required human-decision layer. The comparison is not a CSS filter or static mock: within the judged replay, the preserved live lifecycle receipt is the lost-shade baseline; it completed after 19 bounded polls, spent one event unit, and retained the request and downloaded output. A successful image-bearing baseline is mandatory; if it fails, every decision and outcome remains locked. Candidate fixtures retain outputs, hashes, task ids, and poll counts, but not their requests/lifecycle responses, so LastTube labels them metadata-only rather than calling them receipted. The human must view each usable candidate, explicitly accept or reject it, and prefer one among the accepted set. VTO compares estimated colors consistently; it does not validate the actual formulation or exact variant.

A separate September 3 robustness proof uses a second portrait published in Perfect Corp's API playground. It preserves fresh baseline and candidate-effect lifecycles, sanitized responses, and downloaded output hashes on that different fixture model. It is not substituted into the judged replay and does not establish a product variant, shade, finish, or actionable lead.

For a new live candidate, review stays locked unless LastTube downloads the signed Perfect output and validates a versioned bundle binding sanitized provider-response artifacts, exact listing, source-image bytes, request effects, every task/poll response, safe URL lineage, and output bytes. Offline tests—not a new provider run—prove tampering fails closed. The current recorded candidates predate this path and remain metadata-only; the submission does not upgrade their proof retroactively.

SerpApi: a live Google Shopping search determines which alternatives the response reported and where the evidence came from. A September 3 receipt-only freshness run again normalized 40 candidates without overwriting the deterministic replay. The preserved proof keeps the hero query's merchant, price, exact listing text, observed offer, source, and timestamp fields. Remove SerpApi and LastTube loses its market-evidence claim; remove Perfect Corp and neither the required baseline nor human comparison can occur, so LastTube blocks the outcome.

Challenges we ran into

The hardest boundary was honest fallback behavior. Provider result URLs expire, shopping listings do not prove stock, merchant images are noisy shade evidence, and a delayed live response could once overwrite a freshly selected fixture result. We downloaded VTO outputs, bounded polling, disclosed the data limits, and added a request-sequence guard so stale responses cannot cross truth modes.

Accomplishments that we're proud of

• A complete discontinued-shade-to-candidate-ledger loop with both sponsor technologies load-bearing.
• Sanitized live receipts for a 40-candidate SerpApi response and a successful Perfect Corp VTO lost-shade baseline lifecycle.
• A fail-closed 10% evidence threshold that rejects the preserved 2.5%-coverage fixture candidate.
• No silent omissions: the recorded three-item shortlist resolves into one system exclusion, one human rejection, and one accepted/preferred visual direction.
• Decision-bearing accept/reject/preference controls: the recorded human choice rejects the lower-ΔE item, prefers the other, and still stops because structured action evidence is incomplete.
• Per-candidate offline manifests bind exact listing ids to retained VTO output hashes/bytes and explicitly record every missing input/request/lifecycle field.
• The live server now constructs and exposes a validated, versioned per-run candidate bundle with downloadable sanitized search/lifecycle artifacts and source/output bytes, while failing before provider execution for an unbound candidate and after any retention, cross-field, or tamper error.
• An offline audit of all 40 preserved result titles. Several exact-looking variants exist, but none has stored image bytes needed for a truthful coverage/hash check, so we made zero new VTO calls.
• A deterministic judge-demo route that fails its capture check if it attempts any live provider request or loses its FIXTURE labels.
• 63 offline tests, type checking, linting, production build, secret scan, responsive capture, and a container-ready one-process runtime.

What we learned

Live search and virtual try-on answer different halves of the same decision. Search timestamps the observed market; VTO makes visual judgment possible only when its baseline succeeds. A coverage number without retained source bytes is not provenance, and task/poll metadata is not a lifecycle receipt. Runtime proof also needs an explicit persistence boundary: on the current host, evidence is exportable per run but not durable across redeploy. The useful product is sometimes a trustworthy stop: human preference cannot repair missing evidence.

What's next

Obtain licensed variant-level shade/finish data and retained source-image inputs, then add opt-in availability alerts for evidence-complete exact variants. A disclosed affiliate handoff could fund that workflow, but demand, conversion, willingness to pay, alert delivery, and affiliate enrollment remain unvalidated and unimplemented. Add an explicitly consented user-photo upload only with deletion controls and opt-in studies.

Built With

• docker
• hono
• node.js
• perfect-corp-makeup-vto
• playwright-core
• react
• serpapi-google-shopping
• sharp
• typescript
• vite
• vitest