System

QuoteSnap AI

Applied system: image-assisted estimating where AI interprets and software prices.

QuoteSnap is an applied AI product for property-service businesses: image-assisted estimating, guided workflows, quoting over SMS and email, payments, and opportunity mapping. The research contribution is the split between multimodal interpretation and deterministic pricing. The product contribution is removing slow steps from how those businesses already quote work.

The fact that an LLM can perform a task does not mean an LLM should own that task.
I trust AI to look at a dirty driveway. I do not trust it to decide what that driveway should cost.
A bad result can come from input, context, retrieval, interpretation, schema, tools, rules, memory, or communication.
Demos are not evaluations. Model scores are not system scores.

Problem

Home-service quoting combines information that conventional software handles very differently. Customers describe work inconsistently. Property images contain useful visual signals. Job types vary. Contractors also have their own pricing rules, minimum charges, margins, add-ons, and service-area behaviour.

Research question

Can a production quoting system use multimodal AI to understand ambiguous job information without allowing a probabilistic model to own the contractor's final pricing logic?

Architecture

QuoteSnap separates interpretation from pricing. Customer descriptions, property context, and images enter an AI interpretation layer. The model converts that messy input into structured job signals. Validation sits at the boundary. A deterministic pricing engine then applies contractor configuration to calculate the quote.

QuoteSnap interpretation pipeline

Customer input
  → Job information
  → Property information
  → Uploaded images
  → AI vision and interpretation
  → Structured job signals
  → Validation
  → Deterministic pricing engine
  → Final quote

AI interprets: turn messy, visual, textual, or otherwise ambiguous information into structured signals.
Software decides: apply rules where repeatability, correctness, or business policy matters.
AI communicates: transform the final result into language a human can use.

Fusion, then structure

Visual information
  × User information
  × Property information
  × Application state
  → Structured interpretation
  → Deterministic systems

The resulting pattern is: AI interprets, software decides, AI can communicate. Natural-language generation can still explain the result to the customer, but the economic decision already exists before that communication layer runs.

Why this matters

The architecture creates explicit ownership. AI owns uncertain interpretation. Software owns repeatable business rules. This improves debugging because a wrong quote can be traced to interpretation, validation, contractor configuration, or deterministic pricing rather than treated as one opaque AI failure.

Production system

The product extends beyond quoting into operational workflows such as sending estimates, job and property context, opportunity mapping, and other contractor-facing features. The research value of the system is not that every feature uses AI. It is that the product deliberately limits AI to the parts of the workflow where probabilistic interpretation creates an advantage.

Experiments

EXP-001 documents the rejected model-owned quoting architecture. EXP-010 documents the shift from prose outputs to structured signals at the AI-to-software boundary.

Where a failure can live

Layer	Example failure	Typical fix
Input	Missing photo, ambiguous job text	Collect, constrain, clarify
Context	Wrong files or stale rules in the window	Select, compress, expire
Retrieval	Semantically close, behaviourally wrong	Rank with more signals
Model	Misread surface condition	Examples, schema, specialist model
Schema	Malformed or missing fields	Validate and retry
Tools	Wrong tool or unsafe permission	Restrict, repair, verify
Rules	Bad multiplier or minimum charge	Fix software, not the prompt
Memory	Replaying a discarded decision	Task memory, not raw logs
Latency / cost	Correct but unusable	Cache, smaller model, fewer hops

Changing the model will not fix every AI problem. Layered architectures make the problems easier to identify. The visual system can be scored on interpretation separately from the pricing engine.

What I score separately

Model behaviour on the actual task distribution
Retrieval quality
Structured extraction / schema validity
Tool selection
Task completion
Latency
Cost
Failure recovery
Human usefulness

A general loop

Reliable-AI loop

User / environment
  → Input normalisation
  → Context construction
       ↙ Memory   Retrieval   Application state
  → Model / planner
  → Structured output
  → Validation
  → Tool or deterministic execution
  → Verification
  → State update
  → Final response / next action

Not every product needs every layer. The important idea is separation of responsibility. The model is one component inside a larger intelligent system.

Limitations and artefacts

QuoteSnap is a production product rather than an open research repository. Private customer data, proprietary prompts, and production logs are not published in this archive. The public research contribution is the architecture, the rejected alternatives, and the engineering conclusions derived from operating the system.

Abubakar, Imam. "QuoteSnap AI." Imam Abubakar AI Research, 2026. https://imamabubakar.com/research/quotesnap-ai

This archive contains independent research, experiments, architecture studies, and technical observations. Some findings are exploratory and should not be interpreted as peer-reviewed scientific conclusions unless explicitly stated. Architecture studies of third-party products are based on publicly observable behaviour, published technical information, and engineering analysis. They are not claims of access to private source code or infrastructure.