The architecture, walked end to end.
Six steps from inbound message to cited answer. Three views into what each layer holds.
The assistant operates as a thin layer over an open multilingual foundation model running on the company's own server: retrieval binds inputs to an approved corpus, ingest lint blocks malformed sources, and a post-processing gate enforces citation matching against retrieval output. The same assistant that answers a question logs the interaction — the log is a precondition of citation enforcement, not a feature on top. Hardware and the deployment perimeter live on Controls & Data.
The flow, end to end
Six steps describe what happens between the moment a foreman hits send and the moment the answer lands back in the same thread. Nothing in this sequence is hypothetical — it is the literal path every question takes through the system.
Question arrives in Telegram
A crew member sends a question as text or a voice note to the company's AirOrchestra bot from the jobsite, truck, or break area. No new app to install. Telegram is already on the phone.
Role and job context identified
The user's role and project assignment are pre-set in the company directory. The same question gets a different scope of answer for an apprentice on Project A versus a PM on Project B — because role and project are matched against pre-configured access rules. Roles are not inferred per request.
Knowledge base searched (on the company's own server)
The question is matched against the company's indexed documents: submittals, OEM installation manuals, shop drawings, SMACNA references, ASHRAE tables, coordination drawings, RFIs, and internal SOPs. Index, search, and the model itself live on the same machine inside the company's office. The search looks for sections that directly address the question, not vague topical matches.
Answer built from approved documents (on the same server)
The answer is composed only from passages the system retrieved. Generation runs on the local foundation model — the request never leaves the company's network. Nothing is invented. If the knowledge base does not contain the information, the system says so rather than filling the gap with a guess.
Source citation attached (verified by the local gate)
Every answer carries the document name, section, and page number. A post-processing service on the same machine verifies the cited chunk was actually returned by retrieval this turn — uncited or mismatched payloads are dropped before the user sees them. The citation is not a link dump. It is the specific location a supervisor could open to verify the answer in under a minute.
Answer returns to the same thread
The response lands back in the same Telegram conversation, usually in seconds. The crew member reads it, acts on it, and the exchange is logged. The next question continues in the same thread.
What the employee sees
From the field, AirOrchestra looks like a contact in Telegram. An employee opens the thread, types or records a question, and reads the answer. That is the entire surface area.
The answers are short and carry a citation. A user can double-check any answer in under a minute by opening the cited source. No search engine. No PDF hunting. No phone call to the office interrupting someone else's work.
Illustrative example. The exchange below is constructed to show the shape of a typical answer (citation, document, section). It is not a real production transcript and the specifics shown are not provided as an HVAC reference.
Employee on site
AirOrchestra
Sources cited: Company install standard — Rectangular Duct Hangers (approved by operations lead), Project submittal package — mechanical hanger schedule
What leadership sees
Owners and operations managers do not read every message. They read summaries and flags. Three views describe what the office layer shows.
Cross-project view
A single dashboard lists every active job and the volume of questions coming from each. A job with spiking question traffic usually means coordination issues on site. The view surfaces which projects are running smoothly and which are consuming crew attention.
Risk flags from question patterns
When the same question appears from three different crew members on the same job, the system flags it. Repeated confusion about a detail usually points to a submittal gap, a drawing conflict, or a training need. The flag goes to the PM before it becomes a change order.
Audit trail of every Q&A
Every question, answer, and cited source is logged with timestamp and project tag. At closeout, the record shows what the field asked and what they were told. Useful for warranty disputes, union grievances, and for onboarding the next crew on a similar scope.
The four constraints
AirOrchestra is the same class of AI assistant as ChatGPT or Claude, with four constraints pre-set in advance. Each frame removes a degree of freedom that public AI keeps open by default.
Fixed source
The system answers only from documents loaded by the company. The open internet and general training data are outside the source set.
Citation as gate
Every answer names the document, section, and page. A post-processing service drops responses without a matching citation. Citation is enforcement, not decoration.
Refusal by default
When the approved documents do not cover a question, the system answers "not found in approved documents." No invention, no fallback to general knowledge.
Fixed role and scope
Role and project assignments are pre-set. Access to documents is filtered before retrieval runs. An employee on one project does not retrieve from another, and does not see payroll. Data ownership and export details live on Controls & Data.
Scope of the system
The system has one job: retrieve from an approved corpus and return a cited answer or a refusal. Document authorship belongs to the operations lead. Project tracking, estimating, and accounting belong to the systems already used for that work. The assistant sits next to those layers — it does not replicate them.