How the Nigerian Persona Engine works

Inside the Nigerian Persona Engine — restaurant scoring weights, review tone calibration, multilingual output detection, and the implicit learning loop.

The Persona Engine is the cultural intelligence layer that separates NaijaTaste from a generic restaurant finder. It models how different types of Nigerian diners think, speak, and judge food — and produces output in their language.

What the engine does

Two things, running on every request:

1. Scores restaurants against a user’s persona to produce ranked recommendations
2. Generates reviews in the voice, tone, and language of that persona

Both use the same underlying model — the Nigerian Persona Engine.

How recommendation scoring works

When a user searches, the engine runs a scoring pass:

1. Persona weight vector is loaded for the user’s archetype
2. Restaurant signals are pulled from Google Places (rating, price level, category, review count)
3. Signals are weighted by persona — a Street Food Enthusiast’s score weights price level 3× more than a Lagos Professional’s
4. Proximity decay is applied — restaurants farther from the search area score lower
5. Ranked list is returned, top 5

The whole pass takes under 500ms.

How review generation works

The persona controls four things in every review:

Dimension	What it sets
Tone	How much Pidgin, how formal the language
Star calibration	What a 4-star experience means for this persona
Highlight vocabulary	Which aspects of the restaurant get emphasised
Complaint patterns	What a negative review from this persona focuses on

Multilingual output

The engine produces output in five Nigerian languages based on the user’s channel and preference:

Language	How it’s selected
English	Default on web and API
Nigerian Pidgin	Set by persona or user preference
Yoruba	Detected from WhatsApp message language
Hausa	Detected from WhatsApp message language
Igbo	Detected from WhatsApp message language

On WhatsApp, language detection is automatic — the agent reads your message language and responds in kind. Mid-conversation switches are handled correctly.

The implicit learning layer

On top of the base persona, the engine maintains a learning layer per user. Every interaction is a signal:

save → liked this style
regenerate → didn't like this style
copy → approved the content
adjust → explicit preference captured

The last 50 signals are fetched on each new generation and passed to the model as past_adjustments. The _extract_patterns() function scans for:

• Feedback keywords with frequency thresholds (shorter, less hype, lower rating)
• Saved review rating averages
• Tone save vs. regenerate tallies
• Restaurant type regeneration counts

Matched patterns are appended to the prompt silently. No settings page. No setup.

The three archetypes

The three archetypes

Deep dive on Lagos Professional, Street Food Enthusiast, and The Aunty — their calibrations, vocabularies, and what moves their scores.