AI Wine Purchasing & Palate-Matching Engine

About the Client

A wine business building something ambitious: a single platform that runs B2B wholesale procurement on one side and a personalised consumer sommelier storefront on the other, with the same intelligence layer serving both. Buyers would see a traffic-light buying engine and inventory-gap analysis; consumers would see a match percentage against their palate profile, plus deliberate "palate stretch" recommendations designed to expand taste rather than merely confirm it.

The product vision was theirs and it was sharp. The bottleneck was underneath it: every feature they had designed was, structurally, a data problem they had not yet solved. Match percentages need comparable wine attributes. Traffic-light buying needs live market prices. Gap analysis needs to know what the market offers that the shelf doesn't. Sentiment overlays need reviews at scale. Actowiz Solutions was brought in to build that layer.

The Challenge: Wine Is the Hardest Product Category to Match

Retail entity resolution is a solved-ish problem when products carry UPCs. Wine carries almost nothing reliable, and its identity fragments across five axes simultaneously:

The same wine has a dozen names. One bottle might appear as "Ch. Margaux 2016", "Château Margaux, Margaux 1er Cru 2016", "Margaux 2016 (Bordeaux)", or "CH MARGAUX 16" in a wholesaler's manifest. Producers, cuvées, appellations, and classifications get abbreviated, translated, mis-accented, and truncated differently by every merchant and every distributor.

Vintage is identity, not a variant. The 2016 and 2017 of the same wine are different products with different prices, different critic scores, and different drinking windows — yet merchant listings routinely collapse or omit them. Matching that treats vintage as noise produces price comparisons that are simply wrong.

Format and bottle size fragment the shelf. 750ml, magnum, half-bottle, six-pack case, twelve-pack — normalizing to per-750ml economics is mandatory before any price is comparable.

Attributes that drive matching aren't in the listing. A palate-match engine needs structured attributes — grape composition, region hierarchy, style, body, acidity, tannin, oak, sweetness, drinking window — that merchant pages rarely publish in structured form, if at all. They live in tasting notes, critic reviews, and producer prose.

And the data lives behind aggressive, fragmented surfaces. Wine marketplaces, merchant sites, auction results, and critic databases all change constantly and defend against automation — the classic environment for self-healing extraction.

The client's own summary during scoping: "We can build the algorithm. We cannot build the universe it operates on."

The Actowiz Solution

The wine entity graph. The foundation everything else hangs from. We built a canonical wine identity layer resolving producer → cuvée → appellation → classification → vintage → format as a structured key, with an alias graph mapping the messy real-world strings merchants actually use. Fuzzy matching (string, phonetic, and embedding-based) proposes; validation rules and confidence scores decide; low-confidence matches surface for review rather than silently entering the graph — the flag-don't-guess discipline that governs all our entity work. Manifest ingestion runs through this layer, which is what turned the client's "smart manifest" requirement from a parsing problem into a resolution problem with an answer.
Market price intelligence. Continuous extraction of merchant and marketplace pricing across the client's relevant markets, resolved to the entity graph and normalized to per-750ml economics: current offers, price dispersion per wine-vintage, availability, and historical trend. This is the fuel for the traffic-light buying engine — a wine's buy signal is a function of its offer price against the market's live distribution for that exact wine-vintage, not a static cost-plus rule.
Critic scores and structured attributes. Publicly available critic scores and tasting notes extracted and resolved to the same entity keys, then run through a structuring pass that converts prose tasting language into the typed attribute vectors the matching algorithm actually needs — grape composition, body, acidity, tannin, oak treatment, sweetness, drinking window — with confidence per attribute and provenance per source. This layer is what makes a normalized Euclidean distance meaningful: the distance formula is only as good as the dimensions it measures, and those dimensions had to be manufactured from unstructured text.
Review-sentiment NLP overlays. Public consumer reviews at scale, resolved per wine-vintage, processed into theme and sentiment signals — not a single star average, but decomposed dimensions ("drinking well now" vs "needs time", "over-oaked" complaints, value perception) that the client layers over critic scores. Reviews and critics disagree in commercially interesting ways; surfacing the disagreement was more useful than blending it away. Reviewer identity is masked at the edge and never stored, per our standing PII posture.
Inventory-gap analysis. The client's Shopify catalogue synced against the resolved market universe, producing the gap view their internal dashboard required: what the market is offering (and pricing, and scoring) in the regions, styles, and price bands where their shelf is thin — and, on the flip side, where their shelf holds wines the market has quietly moved away from.
Delivery into the product. Feeds and APIs into the client's stack: entity-resolved catalogue enrichment for the Shopify storefront, attribute vectors into their vector database, price and availability streams into the buying engine, and sentiment overlays into the CRM segmentation logic. Refresh tiered by volatility — prices and availability fast, critic scores and attributes on publication cadence — the delta-based economics from our RAG pipeline work applied to a commerce catalogue.
Compliance and category care. Public catalogue, price, critic, and review data only; PII masked at the edge; per-record lineage; GDPR/DPDP-mapped controls per our compliance framework. One category-specific note: alcohol is a regulated retail category with jurisdiction-specific marketing and sales rules — our layer supplies data, and the client's platform owns age-gating, licensing, and jurisdictional sales logic, which is exactly where those responsibilities belong.

Sample Structures (Illustrative)

Resolved wine record:

{
  "wine_key": "chateau-sample-margaux|grand-vin|margaux|1er-cru|2016|750ml",
  "producer": "Château Sample",
  "appellation": "Margaux",
  "vintage": 2016,
  "format_ml": 750,
  "attributes": {
    "grapes": {"cabernet_sauvignon": 0.85, "merlot": 0.12, "petit_verdot": 0.03},
    "body": 0.82, "acidity": 0.55, "tannin": 0.78, "oak": 0.66, "sweetness": 0.02,
    "drinking_window": [2026, 2045]
  },
  "critics": [{"source": "critic_a", "score": 96}, {"source": "critic_b", "score": 94}],
  "market": {"offers": 14, "median_750ml_eur": 612, "low": 549, "high": 720, "trend_90d": "+4%"},
  "sentiment": {"volume": 218, "net": 0.71, "themes": ["needs time", "classic vintage"]},
  "match_confidence": 0.97,
  "lineage_id": "lin-4471-w"
}

Traffic-light buying view (sample excerpt):

Wine-Vintage (Sample)	Offer Price*	Market Median*	Position*	Critic Avg*	Signal
Sample Bordeaux 2016	€549	€612	−10%	95	🟢 Buy
Sample Barolo 2019	€78	€74	+5%	92	🟡 Hold
Sample Rioja 2018	€41	€33	+24%	88	🔴 Pass
Sample Chablis 2022	€29	€31	−6%	90	🟢 Buy

Sample data — illustrative of deliverable format. Actual feeds are wine-vintage-format level, refreshed on volatility tiers.

Engagement Metrics (Representative)

Metric	Value*
Wines in resolved entity graph	240,000+ wine-vintage-format keys
Alias strings mapped	~1.4 million merchant/manifest variants
Entity-match precision (audited)	96%+ at high-confidence tier
Merchant/market sources tracked	60+
Structured attributes per wine (median)	11 dimensions
Reviews processed into sentiment overlays	3.2 million
Manifest ingestion accuracy (post-resolution)	94% auto-resolved, 6% to review queue
Time to first production feed	6 weeks

Representative engagement figures — illustrative of project structure.

The Outcome

The client shipped both zones of their platform on the data layer, and three effects stood out in their own review.

The match engine got dimensions worth measuring. Their normalized-Euclidean matching logic was sound from day one; what it lacked was a populated, comparable attribute space. Once tasting prose became typed vectors with confidence scores, match percentages stopped being decorative and started being defensible — and "palate stretch" overrides became engineerable, because you can only deliberately move a customer one axis away from their profile if the axes exist.

Traffic-light buying replaced gut-feel procurement. Buyers had been pricing offers against memory and a spreadsheet of last year's costs. Live market distribution per wine-vintage turned that into a signal — and the 🔴 rows (offers priced above the market for wines the market had cooled on) were, per the client, the fastest ROI in the build: the deals they didn't do.

Gap analysis found the shelf's blind spots. The Shopify-versus-market view surfaced whole style-and-price-band pockets where demand signals were strong and their inventory was empty — a merchandising roadmap that had previously been assembled from anecdote and supplier pitches.

The engagement continues as the platform expands: additional markets on the entity graph, auction-result data as a valuation input, and a producer-level layer for the B2B side.

Why This Pattern Repeats in Complex Catalogues

Wine is an extreme case of a general problem: any AI product built on a catalogue is really built on entity resolution. Recommendation engines, price intelligence, gap analysis, and personalization all silently assume that "the same product" can be identified across sources — and in categories without reliable identifiers (wine, spirits, art, collectibles, auto parts, fashion, pharma), that assumption is the whole engineering project. The transferable design: build the identity graph first, manufacture the attributes the algorithm needs from unstructured text, normalize economics (per-750ml, per-serving, per-unit) before comparing anything, and keep confidence scores everywhere so the machine knows what it doesn't know.

Frequently Asked Questions

Why is wine data harder than standard e-commerce data?

No reliable universal identifier, vintage-as-identity, format fragmentation, and the fact that the attributes driving recommendations live in unstructured tasting prose rather than structured fields. Entity resolution isn't a preprocessing step in wine — it's the product's foundation.

Can tasting notes really be turned into structured attributes?

Yes — LLM-based structuring converts critic and producer tasting language into typed dimensions (body, acidity, tannin, oak, sweetness, drinking window, grape composition) with per-attribute confidence and source provenance, which is what makes vector-distance matching meaningful.

How does live market data power a "traffic-light" buying engine?

By resolving every offer to an exact wine-vintage-format key and comparing it against the live distribution of market prices for that same key — so buy/hold/pass signals reflect the market as it is today, not a static cost rule.

Can this approach transfer to spirits, collectibles, or other identifier-poor categories?

Directly — the identity-graph-first architecture is category-agnostic. Contact Actowiz Solutions to scope an entity-resolution and market-data pilot for your catalogue.

Hear It Directly from Our Clients

Watch how businesses like yours are using Actowiz data to drive growth.

▶

1 min

★★★★★

"Actowiz Solutions offered exceptional support with transparency and guidance throughout. Anna and Saga made the process easy for a non-technical user like me. Great service, fair pricing!"

Thomas Galido

Co-Founder / Head of Product at Upright Data Inc.

▶

2 min

★★★★★

"Actowiz delivered impeccable results for our company. Their team ensured data accuracy and on-time delivery. The competitive intelligence completely transformed our pricing strategy."

Iulen Ibanez

CEO / Datacy.es

▶

1:30

★★★★★

"What impressed me most was the speed — we went from requirement to production data in under 48 hours. The API integration was seamless and the support team is always responsive."

Febbin Chacko

-Fin, Small Business Owner