FujiSim GPT

Overview

FujiSim is a custom GPT system designed to recommend film simulation recipes for Fujifilm cameras based on contextual factors like lighting, subject type, and aesthetic direction. The model reasons strictly from a curated CSV dataset (my personal library of film recipes) and returns tailored suggestions grounded in structured metadata rather than generic online advice.

Context

I’ve always loved Fujifilm cameras for their film simulations. They let me fine-tune the color science of each shot to match mood and personality. Over time, I built a collection of custom film simulation recipes that reflect my aesthetic style.

But choosing the right recipe for a given moment isn’t always straightforward. Light changes. Skin tones behave differently. Some simulations thrive in bright sun but fall apart indoors. And because film simulation preference is deeply personal, generic online recommendations often feel too broad or mismatched.

I wanted a solution that understood my personal preferences for film simulations, drawing knowledge from my personal library of film recipes and setting to pick the best recipe to experiment with.

Problem

Film simulation choice is influenced by subtle and highly contextual factors:

• Lighting conditions (tungsten vs. overcast vs. neon)

• Subject type (skin tone, portrait vs. landscape)

• Creative intent (soft, nostalgic, contrasty)

• Desired vibe (warm, cinematic, muted, punchy)

Existing tools don’t account for all of these at once. ChatGPT can guess, but its suggestions come from general knowledge and not my curated recipe set. And manually evaluating recipes every time is slow and inconsistent.

Objective

Design a system that:

• Understands context the way a photographer does

• Recommends only from my personal film simulation catalog

• Uses structured metadata and controlled AI reasoning

• Eliminates hallucination and generic answers

• Produces reliable, explainable guidance for every scenario

Solution

I built FujiSim, a custom GPT that ingests my recipe catalog (CSV) and uses a scoring framework to recommend the best-fitting film simulation for any shooting scenario.

The core design principles were:

• Strict grounding: FujiSim can ONLY use recipes that exist in the CSV

• Structured metadata: each recipe is tagged with lighting, subject type, vibe, and contextual suitability

• Interpretability: every recommendation comes with a transparent breakdown of why it fits

• Personalization: results reflect my taste, not generic Fujifilm advice

This makes FujiSim behave like a personalized color-science assistant.

How it works

1. Structured recipe catalog

Every recipe is annotated with multi-tag metadata, including:

• best_contexts

• light_conditions

• subject_types

• vibe

• detailed settings

• notes; edge cases

These tags use consistent, machine-readable formatting (e.g., portrait_skin_medium).

2. Context understanding

User prompts like:

Indoor tungsten portrait, medium skin tone, warm cinematic vibe

are parsed into 4 dimensions:

• Lighting

• Subject

• Vibe

• Environment

3. Weighted scoring system

Each recipe is scored using:

• Light fit: 0-3

• Vibe fit: 0-2

• Risk penalty: -0 to -1 (e.g., harsh on skin tones)

• Subject fit: 0-2

• Context fit: 0-2

FujiSim ranks the top matches and explains the reasoning behind each.

4. Strict knowledge boundaries

To prevent hallucination:

• FujiSim MUST load the CSV before answering

• It cannot recommend anything outside the catalog

• If a requested sim doesn’t exist, it returns nearest matches

This creates predictable, controlled behavior — crucial for an AI tool working with personal taste.

What I learned

1. Structured reasoning with AI: Explored how to translate subjective creative preferences into structured metadata that an AI can interpret consistently.

2. Designing controlled outputs: Preventing hallucination taught me how to set strict boundaries for LLMs, something highly relevant in AI system design.

3. Human–computer interaction for creative tools: FujiSim sits at an interesting intersection of aesthetic intuition and technical logic, aligns closely with my interest in emerging technologies.

4. Building personalized AI assistants: I learned how to create tools that adapt to personal workflows rather than broad generic use cases.

Final thoughts

FujiSim started as a personal experiment to make my photography process smoother. It grew into a full demonstration of how AI can reason over structured datasets and support creative decision-making without losing personal nuance. The experience building and using FujiSim reminded me that I enjoy building tools that support my creative pursuits and not just what’s technically possible.