All cases
Case 04 · iOS · Vertical AI · Production

Sommvi.

An AI wine sommelier that surprised
a championship judge.

Built and operated solo. Live on the App Store →

How agents run the product, and the team
Scene 01 · The thesis

Two systems. One product.

Sommvi is unusual because there are agents on both sides of it. There’s an agentic system inside the app — multiple specialist agents that compose every wine answer. And there’s an agentic team around the app — Tech Lead and Product Owner agents that build, deploy, monitor, and grow it.

That dual-angle is what makes this case load-bearing for the playbook: agents aren’t just inside your product or just around it. They’re both. Same patterns, different scopes.

“Intrigued and surprised by what the app produced — I didn’t expect this kind of pairing intuition from software.”

Senior judge · International wine-tasting championship

INSIDE THE PRODUCT The agents that answer. USER · WINE Q&A "What pairs with…?" COORDINATOR Routes & composes KNOWLEDGE 50K wines REASONING Pairing logic SYNTHESIS Voice + format EXPERT ANSWER Barolo · 94/100 AROUND THE PRODUCT The team that ships it. TL · SRE Tech Lead + orchestrates 2 engineers PO · MARKETING Product Owner Roadmap · copy · growth ENG A code ENG B review PEER REVIEW SOMMVI The product SAME PATTERNS · DIFFERENT SCOPES
Scene 02 · Inside Sommvi

A query becomes a defensible answer in 1.2 seconds.

The user-facing experience is intentionally simple — natural language in, expert response out. The complexity sits behind the interface, in a four-stage agentic pipeline.

A coordinator routes the request to specialist agents that work in parallel: knowledge lookup, pairing reasoning, taste memory, and synthesis. The pipeline composes a structured answer with a confidence score — not generic LLM output.

USER QUERY · STAGE 0 "Pair with osso buco?" COORDINATOR · STAGE 1 Route & orchestrate KNOWLEDGE Wine DB 50K bottles REASONING Pairing logic Multi-step SYNTHESIS Voice + format Sommelier tone RESPONSE · STAGE 3 · 1.2S Barolo, 2018 CONFIDENCE 94 / 100
Scene 03 · The 2nd brain

A knowledge graph, not a prompt.

Sommvi’s answers don’t come from a single LLM call. They come from a structured knowledge graph that connects wine ontology, pairing patterns, regional context, vintage data, and user history — composed at query time.

The 2nd brain is the unfair advantage. Generic LLMs hallucinate vintages and pair Bordeaux with sushi. Sommvi grounds every claim in the graph, then lets the language model do what it’s good at: turning structured facts into a sommelier’s voice.

QUERY composes REGION Piedmont… VARIETAL Nebbiolo PAIRING Rich meats PROFILE Tannic HISTORY User taste VINTAGE 2018, hot PRODUCER Mascarello 7 DOMAINS · ~50K NODES · ~120K EDGES
Scene 04 · Around Sommvi

Two seniors orchestrating two builders.

The team is four agents. Two senior agents — Tech Lead and Product Owner — own the strategy. Each doubles up: TL also handles SRE; PO also handles Marketing.

Under TL, two engineer agents do the actual building. They code in parallel and peer-review each other before anything ships. The TL orchestrates them, reviews the final output, and owns reliability. Real software-team patterns at single-operator cost.

SENIOR · 2 HATS Tech Lead DOUBLES AS SRE SENIOR · 2 HATS Product Owner DOUBLES AS MARKETING ORCHESTRATES 2 engineer agents ↓ SCOPE ROADMAP · COPY · GROWTH (single agent) ENGINEER A Builder Code · test Review B ENGINEER B Builder Code · test Review A ⇆ PEER REVIEW EQUIVALENT TRADITIONAL PLAN 10 PEOPLE · 6 MO TL · 4 ENG · PO · MKT · DESIGN · 2 SPEC SOLO + AGENTS · 4 AGENTS <30 DAYS · LIVE £300/MO RUNTIME
Scene 05 · The compression

Same scope. 30 days, not 6 months.

The traditional plan for Sommvi’s scope was sized for 10 people over 6 months: a Tech Lead, four engineers, a Product Owner, a marketing lead, a designer, two domain specialists. That’s the proposal you’d normally take to a venture firm.

We shipped solo with agents in under 30 days. Live on the App Store, validated by a senior judge from a wine-tasting championship. Stack runs at £300/month. The headline isn’t cost-savings against a team that never existed — it’s a 6× compression in time-to-live, with 10-person scope absorbed by one operator and four agents.

6 MO 4.5 MO 3 MO 1.5 MO 0 6 MO · 10 PEOPLE TRADITIONAL PLAN TL · 4 ENG DESIGNER PO · MARKETING 2 DOMAIN SPEC. <30 DAYS SOLO · LIVE SOLO + AGENTS 6× FASTER CONCEPT → LIVE · SAME SCOPE
Next More cases · or work with me

Different business. Same playbook.