An HVAC AI assistant with four architectural constraints wired in.
Fixed role. Fixed source. Fixed scope. Citation or refusal. An open multilingual foundation model running on the company's own server — with three enforcement layers around it.
Built and run inside one commercial HVAC company. The model and the company's documents share one perimeter — a single workstation in the company's own office. This site documents the system end to end — retrieval, ingest lint, post-processing gate, channel layer — so an operation that wants to build something similar can replicate it.
Your records. Your server. Your lock on the door. The corpus does not leave the building.
What the architecture is built to do
Turn every question a crew member asks during the working day into a record that stays with the company when the person who asked it does not.
The journeyman who knows which fitting goes where on a downtown high-rise from the 1960s carries that knowledge in his head. When he retires, decades of installs, every callback, every “we tried this once and it failed” — go with him. The company that hired him at twenty-two and paid him for forty years of judgment ends up, the day after his retirement party, marginally smarter than it was before he started.
The architecture moves the locus of that knowledge from the individual to the indexed corpus. Retrieval can surface a 2026 install detail in 2036 for the next apprentice on a similar job. The company gets older, the corpus gets deeper, and each new hire walks into more institutional memory than the previous one — instead of less.
The trade-off, named: the corpus has to be populated, and populating a corpus is staffed work. A company with sloppy closeout discipline produces a thin retrieval no matter what the architecture promises. The architecture amplifies the closeout discipline the company already has — it does not create one.
Two paths to AI inside an HVAC company
Train people to use public AI
Discipline lives on the user. Fifty employees write fifty different prompts. Quality and information hygiene depend on whoever sat through the training. Even disciplined senior officials have leaked sensitive material into public chat.
Build the architecture in-house
Discipline lives in the system. Retrieval binds to an approved corpus, ingest lint blocks malformed sources, a post-processing gate drops uncited responses. The user does not have to remember to behave.
The first path is what most companies are running today. The second is what this site documents.
The four constraints
Each constraint removes one degree of freedom that public chat keeps open by default.
Fixed role
The assistant operates as an HVAC specialist tied to the user's role in the company. Off-scope topics return "out of scope for this assistant."
Fixed source
Answers are built only from approved company documents. Retrieval physically pulls from that corpus and only that corpus — not from a system-prompt instruction.
Fixed scope
Document access is filtered by user role and project assignment before retrieval runs. Roles are pre-set in the company directory, not inferred from how the question is phrased.
Citation or refusal
A post-processing service runs between model output and reply send. It verifies the response cites a chunk that retrieval actually returned for this turn. Uncited or mismatched responses are dropped — the user receives the refusal string instead.
What approved documents look like
Six categories of source. Industry codes, OEM documentation, company standards, project history, project specifications, safety documentation. Nothing else.
Industry codes & standards
SMACNA duct construction. ASHRAE 90.1 and 62.1. NFPA 90A. IBC and IMC mechanical chapters. NEC sections relevant to HVAC. OSHA construction standards.
OEM documentation
Equipment manuals, startup procedures, fault codes, wiring diagrams, commissioning sequences for the manufacturers and models the company actually installs — Carrier, Daikin, Trane, Mitsubishi, LG. Not generic internet specs.
Company standards & SOPs
Approved install methods, hanger schedules, quality checkpoints, accepted field practices. The way the company does things, written by operations and signed off by ownership.
Project history & specs
RFIs, change orders, lessons learned from closed jobs, contract documents and approved shop drawings for active work, site-specific safety plans. Closeouts feed the corpus — the next bid sits on the company's documented experience.
Why the audit log exists
The system cannot enforce citation-or-refusal without recording what was retrieved. Every interaction stores user, timestamp, project tag, retrieved chunks, response text, and citation. The log is a precondition of the four constraints — not a feature on top of them. A side-effect: when something contested surfaces later, the record of what was asked and what was answered is already there because it had to be.
Where to go from here
Three entry points into the documentation. The approach names the four constraints and why each one matters. The architecture walks the system end to end — retrieval, lint, post-processing, channel layer. The trade-off log lists what was deliberately not built and why. If you operate a similar HVAC company and want to compare notes on what worked and what failed, write.