Kenny's Thinking Garden

Food Recommender

6 min read

Food Recommender

Core Idea

A food recommendation platform — covering hawker centres, cafes, and artisan spots in Singapore — that normalises recommendations against the user’s own taste preferences, rather than treating a reviewer’s opinion as universally applicable.

State of Thinking

The core problem is a mismatch between reviewer preferences and reader preferences. When someone follows a food recommendation, they are implicitly trusting that their tastes align with the reviewer’s — which is often not true. The platform solves this by profiling both sides and bridging the gap.

Critical constraint — the reviewer profile problem:

The core mechanic only works with profiled critics, not crowd-sourced reviews. If users become the reviewers, the differentiation collapses — that’s just a standard recommender system (Burpple, Google Maps, collaborative filtering). The product is only distinct if reviewer profiles are credible, pre-established, and meaningfully different from each other.

This means the path to reviewer profiles matters enormously:

  • Scraping reviewer sites is legally risky and operationally fragile (TOS violations, per-site parsers, stale data)
  • Asking reviewers to self-profile is unlikely to scale
  • Most promising path: AI inference — use published writing from food critics to infer palate traits, rather than scraping raw scores. This could be the real technical moat of the product.

MVP approach — anonymous reviewer archetypes:

Using real named reviewers carries legal risk (defamation, personality rights, false endorsement). For MVP, reviewer profiles are replaced with three fictional archetypes that cover the Singapore food landscape. Users pick the archetype whose palate most resembles reviewers they trust, and normalisation happens against that archetype’s profile.

The three archetypes:

  • The Hardcore Local — hawker-centric, high spice tolerance, strong wok hei preference, values authenticity and tradition. Benchmark dishes: BCM, Char Kway Teow. Suspicious of anything fusion or tourist-facing.

  • The Weekend Foodie — balanced palate, values consistency and value-for-money. Eats at mall restaurants and family-friendly places. Occasional sit-down restaurant for occasions. The Burpple-sharing, family WhatsApp type.

  • The Craft Chaser — seeks intentionality in food and drink. Regulars at specialty coffee spots, queues for a specific croissant, notices when bread is properly fermented. Equally at home at a third-wave coffee bar or a traditional kopitiam, as long as someone clearly cared. Searches: sourdough, single origin coffee, Basque cheesecake.

Three archetypes covers the Singapore food landscape cleanly — hawker, family/mall, and cafe/artisan — without overlapping or overwhelming the user during onboarding.

The problem:

  • Every person has different taste preferences and tolerance levels — what is mildly spicy to one person is unbearable to another
  • Food influencers and reviewers give recommendations from their own palate, with no signal to the reader about how their tastes differ
  • Readers currently have to manually normalise recommendations themselves — an invisible, effortful step that most skip

The solution approach:

  • Profile users through an initial set of questions to establish their taste preferences
  • Profile reviewers and influencers the same way — create a taste fingerprint for their recommendations
  • Use the gap between the two profiles to contextualise recommendations — e.g. “this reviewer rates spice tolerance 4/5, you rate 2/5, adjust expectations accordingly”
  • Improve accuracy over time through post-visit feedback loops

Feedback loop mechanism:

  • After a visit, prompt the user with specific verification questions — e.g. “Based on your visit, do you think this restaurant is child-friendly?”
  • Real experience feedback refines both the recommendation accuracy and the user’s preference profile over time

Two-sided profiling:

  • User side: establish preferences upfront, refine through feedback
  • Reviewer side: extract preference signals from their existing reviews and ratings, attach as caveats or labels to their recommendations

Thinking Trace

  • The mala soup example made this concrete — “least spicy” means completely different things to different people
  • The Japan ramen example showed this has been partially solved before in a physical context — ramen shops grouping stores by noodle hardness, broth colour etc. on a multi-point scale
  • The gap is that online recommendation platforms do not do this — they treat all recommendations as equivalent regardless of who is giving them
  • The interesting design challenge is getting preference data from reviewers passively — they are unlikely to fill in a profile form, so it has to be inferred from their existing content
  • Post-visit feedback as a refinement loop feels low friction if the questions are specific and quick

What changed?

v3

Two decisions: (1) product scope broadened from hawker-only to include cafes and artisan spots — driven by the Craft Chaser archetype requiring a wider context. (2) Named reviewers replaced with three fictional archetypes to eliminate legal risk at MVP stage — The Hardcore Local, The Weekend Foodie, The Craft Chaser.

v2

Model shift: identified that the user-as-reviewer path kills the core differentiator. The product is only distinct if reviewer profiles are credible and pre-established. Clarified that AI inference from published critic writing is the most promising path to building those profiles.

v1

— skip —

Why it changed?

v2

Prompted by scoping a Lovable prototype — specifically the question of where reviewer data comes from. Working through the options (scraping, partnerships, user-generated) surfaced that user-generated reviews make PalatMatch indistinguishable from existing platforms. The AI inference angle emerged as a potential moat worth tracking.

v1

— skip —

Evidence / Signals

  • The Japan ramen categorisation system is real-world proof that taste dimension profiling works and is valued by users
  • The problem is common enough to be a shared frustration — most people have had the experience of a disappointing recommendation

Open Questions

  • How do you build reviewer preference profiles passively — inferred from their content rather than requiring them to fill in a form?
  • What is the minimum viable preference profile — how many dimensions are needed before recommendations become meaningfully personalised?
  • How do you handle reviewers who cover a wide range of cuisines vs specialists?
  • What does the user experience look like — does the user search for a restaurant, or does the platform push recommendations?
  • How do you bootstrap the reviewer side — do you seed it manually or wait for organic reviewer sign-ups?
  • Is Singapore a good initial market for this given the density and diversity of food options?
  • Can AI reliably infer palate traits from food writing with enough consistency to build profiles?
  • Which Singapore food critics would be viable starting reviewer profiles?
  • Is the AI inference angle defensible as a moat, or easily replicated?

Connections

  • (none yet — first node in this tree)

Changelog

  • v3 — Scope broadened to cafes and artisan spots; named reviewers replaced with three fictional archetypes for legal safety at MVP stage
  • v2 — Model shift: user-as-reviewer path kills differentiation; AI inference from critic writing identified as most promising reviewer profile approach
  • v1 — Initial node

Graph View