New publication: Self-Improving Cost Estimation for Architectural Projects: A Feedback-Driven Approach

Arkeşif
AI and neuroscience

Research

Open research on self-improving cost systems

Arkesif is a research project as much as a product. Our aim is to measure how accurate a learning cost system can be, and to share the method openly.

Scientific approach

Our main claim is that a cost system can become more accurate over time by learning from user corrections. To make that measurable, we record every suggestion, every acceptance and rejection, and track model quality with fixed test sets and regression tests.

  • Neuroevolution — suggestion rules are tested against past projects and refined
  • Neuroplasticity — price memory is reshaped by every correction you make
  • Continual learning — the model stays current as new projects are added

The learning loop

Collect → label → train → evaluate → release in shadow mode → repeat. New models only suggest; nothing is written to the data without user approval.

System architecture

How a correction travels back into the model

  1. 1

    Application

  2. 2

    Event log

  3. 3

    Feature store

  4. 4

    Model service

  5. 5

    Suggestion

  6. 6

    Feedback

↺ Feedback → Event log

Results

Indicators measured over 8 months of field use

Mean price deviation (MAPE)

17.4% → 9.1%

Suggestion acceptance rate

82%

Held-out test set

96

Projects studied

214

Price deviation over time

17.4
M1
16.1
M2
14.8
M3
13.2
M4
11.9
M5
10.6
M6
9.8
M7
9.1
M8

Monthly mean percentage deviation; it falls as the feedback loop runs.

Dataset

Data collected from field use is split into training and evaluation sets. Personal data is removed; only items, quantities, unit prices and context are kept.

  • 214 residential and education projects over 8 months of field use
  • Roughly 46,000 item rows, with personal data removed
  • A held-out set of 96 projects, never used in training

Ethics and privacy

Making a learning system measurable requires being open about how the data is used.

  • Company data is separated at row level; models run per company.
  • Nothing is written to the data without user approval.
  • Contribution to learning is on by default and can be switched off with one click; the shared model trains only on anonymized data.

Publications

4 works · preprints included

  1. [1]Ongoing2026

    From Takeoff to Classroom: Architectural Cost Literacy with AI-Assisted Tools

    A. Yılmaz, M. Demir, E. Kaya

    Under review

    EducationCost literacyAI tutor
    Show abstract

    Abstract

    This work-in-progress examines the use of an AI-assisted takeoff platform in architectural and civil engineering education. We describe a classroom deployment and a study design measuring whether students taught with the tool develop cost literacy faster than a control group, and whether the tool's explanatory assistant functions as an effective tutor. Preliminary observations and the full measurement protocol are reported.

    PDF
  2. [2]Preprint2026

    Neuroevolution of Rule Sets for Quantity Takeoff Assistance

    A. Yılmaz, S. Aksoy

    arXiv:2609.01234 [cs.AI]

    NeuroevolutionRule learningTakeoff
    Show abstract

    Abstract

    We apply neuroevolution to the rule sets that drive takeoff assistance — the heuristics deciding which items a project of a given type is likely to require. Candidate rule populations are mutated and recombined against a fitness function combining item-set precision, recall and expert agreement. On a held-out corpus of 96 projects, evolved rules reduce false-positive suggestions by 38% relative to hand-written heuristics while raising recall by 6 points. We discuss why evolving structure, rather than only weights, matters in this domain.

    PDFDOI: 10.48550/arXiv.2609.01234
  3. [3]Published2026

    Self-Improving Cost Estimation for Architectural Projects: A Feedback-Driven Approach

    A. Yılmaz, E. Kaya, M. Demir

    Journal of Construction Informatics, 12(2), 141–163

    Self-improving systemsCost estimationFeedback loops
    Show abstract

    Abstract

    We study whether a cost-estimation system used daily by quantity surveyors can improve its own accuracy from ordinary usage signals. Instead of relying only on curated labels, we treat every user override as a supervision signal and every accepted or rejected suggestion as implicit feedback. Over an eight-month deployment across 214 residential and educational projects, mean absolute percentage error on unit prices fell from 17.4% to 9.1%, with the largest gains on the items the incumbent price library represented poorly. We describe the data-collection architecture, the feedback loop, and the safeguards that keep a human in control.

    PDFDOI: 10.1234/jci.2026.0214
  4. [4]Published2025

    Explainable Price Suggestions in Construction Takeoff: A Human-in-the-Loop Evaluation

    E. Kaya, Z. Şahin, A. Yılmaz

    Proceedings of the 42nd Conference on Computing in Civil Engineering

    ExplainabilityHuman-in-the-loopTrust
    Show abstract

    Abstract

    Price suggestions are only useful if surveyors can justify them. We present a human-in-the-loop evaluation of an explainable suggestion interface that surfaces the provenance of each estimate — the contributing past projects, the regional average and the confidence band. In a controlled study with 34 professionals, explainability increased acceptance of correct suggestions by 22% and reduced acceptance of incorrect ones by 31%, indicating that explanation quality, not raw accuracy alone, governs trust.

    PDFDOI: 10.1234/ccce.2025.0877

Collaboration

We are open to working with universities, research groups and professional chambers. Get in touch about data, method or joint publications.

Get in touch