Useful, manageable, direct improvements, built in reusable parts so established work can adapt as new needs appear.
the data is never finished
Practice first
AI development at KSR has grown from BIM and visual workflows into documents, knowledge, marketing, branding and commercial applications. Working in a small team, alongside day-to-day BIM responsibilities, has shaped an approach focused on useful, manageable and direct improvements. Systems are designed in reusable parts, so established work can be adapted as new needs emerge.
Method
Start by understanding the task, the people carrying it out and the information it depends on. Develop a working response, test it against actual requirements and use feedback to refine it. Building tools often exposes gaps in the data or instructions behind them, so development improves the wider process as well as the tool itself.
Current focus
With visual workflows established, development now concentrates on agents, practice data and connections between tools. Clear interfaces, reusable instructions and structured references make complex processes easier to access. Human review remains a vital part of agentic workflows wherever professional judgement is required.
Practice data into tools
Fee Calculator
Two versions: base and comprehensive
In useFurther versions planned
The Fee Calculator illustrates this approach. Existing Union Square records provided a starting point for a project library and a benchmarking interface. As the application developed, it exposed information that needed completing or amending. The same work now supports wider uses of the project information, including marketing applications and the KSR website refresh.
Project library | OmniChat prompt, agentic workflow, Aug 2025
Benchmarking view | Compare a proposal with selected historical projects
Data that develops through use
Information gathered for one task can support others. As it is reviewed, enriched and reused, the gaps in the data become clear and new requirements push for additional datasets. Each application improves the shared information available for the next, so the dataset develops alongside the practice’s needs.
01
Gather & compile
An AI-assisted extraction workflow organises CRM exports and supporting PDFs and Word documents into consistent project records. It began with Excel exports and later extended to other formats. Structured records made missing fields easier to spot, and extra information was collected where needed.
02
Apply & refine
The records feed the calculator’s searchable project library and benchmarking views. Testing exposed missing or inaccurate areas and fees, which fed back into the records. Both versions offer a practical way to explore previous projects alongside a current proposal, and they are only as useful as the data underneath.
03
Reuse & extend
Marketing applications are adding consultant information and distinctive project features. A separate prototype explores a searchable record of project issues and lessons learned, linking each problem to its context, response and references, so experience is easier to retrieve when a similar question comes up on another project.
Current position & next steps
Planned refinements include adjusting historical fees for inflation and improving comparisons through project complexity and distinctive features, such as sorting by projects with a swimming pool or spa. The data is not static; it is built over time.
From families into agentic tools
Door & wardrobe libraries
The door and wardrobe libraries show how technical work can lead to useful software. Building the families exposed recurring production tasks that could be automated, while conversations with colleagues revealed ways to make the finished libraries easier to use. Development has grown from maintaining components to supporting the people who configure and use them every day.
Door showroom | Revit library | Present
Building the foundation
As BIM Manager, I developed KSR’s internal door and wardrobe family suites around project requirements, using nested components such as panels, handles and architraves that combine across different configurations.
This gives flexibility and helps with schedules and tagging, but introduces relationships that need careful management. Dimensions must control the intended geometry, components must respond together and parameters must stay understandable as the library grows.
Developing supporting tools
AI-assisted coding workflows produced pyRevit utilities around these tasks: testing dimensional changes, inspecting parameter use and linking controls between nested components and the host family. Modelling knowledge provided the basis for reviewing the generated code and checking its behaviour in Revit.
01Link
Link parameters between nested components and the main family for coordinated control.
02Flex
Change parameter values to test how a family responds and find problems in its behaviour.
03Param
Inspect which parameters a family uses, to support review and further editing.
Door Builder
A custom family builder using premade components
In developmentDoors & wardrobe families
Colleagues highlighted a different need. The nested structure offered plenty of flexibility, but configuring it required confidence with components and parameter relationships.
That prompted Door Builder: an interface for selecting and replacing components within existing and new families. Coded in pyRevit through agentic workflows, it brings together available replacements, visual previews and parameter compatibility feedback, so users understand a change before creating a variant.
Testing continues on component replacement and the behaviour of resulting families, with wardrobes next: a similar but quite different set of nested elements.
Door Builder | Family configuration | In development version
Select the panel, handle or other component to replace
Compare the replacement and the family previews
Create a variant and set the new family name and output location
User feedback shaped development
The first utilities supported the technical work of building and maintaining the library. Colleague feedback turned attention to the experience of using it, which changed the priority: make the flexibility already in the families accessible through a clearer interface.
Current position & next steps
The family suites are in use; the builder is being tested before rollout to a small group of colleagues. Testing focuses on component replacement, parameter behaviour and common configurations. Functional stability comes before any further work on the UI.
Agent development & integration
Product information to Revit families
In developmentBIM content
Handles were previously modelled one by one as requests came in from the interiors team. This in-development workflow connects product research with family creation: one agent retrieves and organises technical information, a second uses reviewed references and detailed modelling instructions to build draft Revit families and test them against set goals and criteria. Human review sits between stages and before anything enters the practice library.
01
Retrieve & review information
A retrieval agent finds product pages, drawings, images, dimensions and technical references suitable for family creation. The collection approach varies with the source and the material available.
Testing narrowed collection to files that support modelling. That made the output easier to review and exposed missing dimensions or unclear component relationships before generation began.
Retrieval agent | Source pack for a door knob
Human review 01 / Source information
Check the references are relevant, sufficiently detailed and consistent. Flag anything that needs clarifying before progressing.
02
Apply modelling knowledge
Initial tests with ChatGPT’s Astra model explored what could be generated from product information with little guidance. The results showed where the practice’s family-modelling approach needed translating into clear rules for the agent.
Reference files now cover family structure, reference planes, naming conventions and parameter relationships: technical knowledge that otherwise stays implicit when modelling by hand. The instructions develop alongside the tests, and each problem in a generated family is evidence of what needs explaining or checking next time.
Handle collection | Reference sheet and family types
03
Test & expand
The first batch produced 20 draft families on request. Review found components moving independently, alignment issues and dimensional parameters failing to control geometry consistently.
So a family’s initial appearance is only part of the assessment. Testing also checks how it responds to dimensional changes and whether relationships between components behave as intended.
04
Refine instructions
Findings fed revisions to the modelling guidance and more specific checks for the agent. For multi-part handles this meant clearer component relationships and checks that geometry stays coordinated when dimensions change.
The loop connects each observed problem to a change in the instructions, and the agent takes part in developing its own checks.
Agent-built families | Parameter tests | In development
Human review 02 / Family behaviour
Check dimensions, geometry, component movement, parameter control and hosting where applicable.
Current position & next steps
The first batch remains a development test. Next: repair and verify a different set of handles against the refined instructions, then repeat the first batch. Further work would extend the approach to other product types. Tests on complex window families have been very promising.
AI productivity templates
Transcript to meeting minutes
In useDocument templates
Reusable AI templates support research, design, documentation and marketing. The Minute Maker is one of the office’s most used workflows, shaped by colleague feedback and by advances in AI model capabilities.
01
Copy & paste into Word Online
Generated minutes were first pasted into Word Online, where markdown formatting worked, unlike in the desktop app. It still needed manual formatting, as the result stayed slightly unstructured and unevenly pasted.
02
Custom MS Word macro
Keeping the existing minutes template as intact as possible became a requirement, so a custom Word macro transferred the generated content into that familiar format.
03
Direct template editing
Once AI tools could work directly with Word documents, the macro became unnecessary. The template and markdown instruction files now guide an agentic workflow that populates the template straight from the raw transcript.
Minute Maker
Raw transcripts are combined with the existing Word template and markdown instructions. Project-specific dictionaries help interpret names, company abbreviations and shorthand that AI might otherwise misread. The instructions define how discussions, decisions and actions are organised.
System prompt and example transcript | Placeholder data
Meeting minutes, KSR template | Placeholder data
Human review 01 / Before issue
Check names, decisions, action owners and unresolved points against the transcript.
Part of a wider template library
The same approach supports other recurring agentic workflows and agents, with reusable instructions, reference files and consistent outputs: research helpers, material finders, visualisation prompt engines, project summary writers, marketing and website writers, DAS and text editors, portfolio creators, planning application reviews and more.
Template library | OmniChat agents
Current position & next steps
The workflow now populates the KSR Word template directly. Refinement focuses on reference files for project-specific terminology, output consistency and feedback from use. The tool is packaged for use by an agent.
All project data and metrics have been anonymised to maintain client confidentiality.