R&D Center Update
This R&D update describes testing of structured detection and fact extraction features for KK Field Logger on real site photos. The goal is to support project records, reviews, and checks with traceable outputs that still require human oversight.
Evidence AI Stage
Published
May 9, 2026
Category
Company Update / R&D Center
Product Direction
KK Field Logger
Scope Note
This article describes current R&D validation and platform direction. AI outputs are designed to support review, not replace project managers, finance reviewers, inspectors, or professional judgment.
K&K Data Service has run a validation round for new AI features added to KK Field Logger. The tests cover structured detection and fact extraction on top of the existing upload and record tools.
Previously, field photos were typically stored as files. The tests explore making each photo, video, voice note, receipt, GPS point, and note part of a traceable record that can be reviewed and summarized.
KK Field Logger already supports field uploads from app and H5, photo and video records, receipts, voice comments, GPS points, project views, permissions, comments, notes, maps, recovery, batch review, and reports.
AI queue processing, receipt fact extraction, and related report features are in active R&D testing. These add structured data on top of the media records.
Ordinary captions like "people are working on site" give little detail for records. Tests focus on whether outputs can note how many people appear, visible activity, materials or tools, and location cues that might support a record.
The tested approach favors short structured fact cards, confidence scores, and explicit conflict flags over free-form text.
Current tests organize processing in layers: a detection step identifies people, objects, and location cues. A second step generates a short structured fact card. When the two steps disagree or confidence is low, the output is flagged for review instead of producing a single answer.
Summaries for different roles draw from the same flagged fact layer. Reviewers are expected to check the source observations.
Validation used 20 real project photos. These show framing work, material handling, staging, vehicles, tools, foundations, distance views, partial occlusion, and varied angles.
Fact card and detector matched on 14 of 20. In harder cases with occlusion, moving crews, or distant areas, outputs were flagged for uncertainty rather than forced to a single result. All flagged items require human review before use in records.
Tests include structured indicators for visible people count, activity type, materials, tools, PPE presence, site state, and basic matching of receipts or GPS to the project record.
These are test outputs only. A reviewer must still inspect the original photos and decide what belongs in any record or report.
Tests used actual photos from KK Field Logger uploads. Real site images include backlighting, motion blur, distance, clutter, partial views, and changing conditions. These factors are why outputs carry confidence scores and review flags.
AI outputs do not replace project managers or reviewers. A site can generate many photos daily. The tested features flag items for review and surface candidate facts so that a human reviewer can locate details faster.
Any AI suggestion must be checked against the source photo, video, or note before it is used in a record, report, or decision.
Tests include options to run vision and language models locally so that photos and data can stay inside a controlled environment.
Current test features cover local model access, job queues, report history, basic query support, and receipt extraction.
KK Field Logger field capture and the current AI review tests are developed by K&K Data Service Inc. Work covers the data models, permission rules, evidence processes, layered analysis, conflict flagging, and report generation used in the tests.
The system uses licensed open source components and local AI runtimes as needed. K&K reviews applicable licenses during development.
Integration of capture, flagging, and review views is the part developed for this platform.
Next steps under test include expanding the set of real field photos used for checks, improving detection of people, materials, and equipment, storing flagged observations for later review, and having query features draw from the verified fact layer.
These remain R&D tests. No outputs are presented as final records without human review.
We are not trying to build an AI that writes essays about photos. We are building a system that helps companies understand field conditions, trace evidence, reduce missed details, and support management decisions.
Current tests add review flags to uploads so observations can be checked against the source media.
