A1transcribe
Field transcription
Streaming speech-to-text tuned on HVAC vocabulary, refrigerant codes, part numbers, and brand model schemas. Handles rooftop noise, PPE muffling, multiple speakers.
WER 3.1% · HVAC eval
Agents for hvac operations.
The agent layer for the service economy. Let your team do the field work and Groundline handle everything around it.
The problem
Service is a chain of handoffs. Closeout is where it breaks.
Field
Technician admin drag
Closeout runs five to forty minutes a job. In noisy conditions, under time pressure, typed notes degrade fast. Photos get taken but never labeled. Equipment IDs get guessed at later.
Office
Rework and follow-up
When closeout is incomplete, dispatch and admin rewrite notes, chase missing photos, and hold invoices. It's not one-off friction. It's a recurring cost on every job.
Platform
Standardization risk
Every branch brings its own workflow, forms, and closeout cadence. Branch KPIs become non-comparable. Shared-service consolidation stalls. Roll-ups fail to roll up.
§ 02 · Product
A multi-agent system,
not a single model.
Five specialist agents coordinate on every job. Each is independently evaluated, versioned, and replaceable. We publish an internal leaderboard against a real-service eval suite, with no fixed demo runs.
A2vision
Equipment vision
Reads nameplates, extracts model and serial numbers, identifies components. OCR + structured VLM pipeline trained on 140k+ nameplate images across major manufacturers.
98.4% serial accuracy
A3orchestrate
Orchestration
The planner. Decides which specialist runs when, what's missing, when to ask the tech a clarifying question. Built on a controller model with tool-use over the full agent fleet.
Typed plan graphs · replay
A4schema
Schema extraction
Maps unstructured output to the branch's exact closeout schema, not a generic template. Learns the shape of your FSM fields, validation rules, and required attachments.
Per-tenant fine-tune
A5verify
Completeness & QA
First-pass checks flag missing photos, inconsistent serials, unlikely labor entries. Fires while the tech is still on site, not after the admin finds it three days later.
99.1% catch rate · 0.6% false-flag
↗handoff
FSM integration
Posts an invoice-ready packet to ServiceTitan, Simpro, COINS, or your internal system. Drafts follow-up quotes. Human reviews before anything reaches the customer.
Native adapters · no middleware
The evidence
The mechanism is already proven in the field.
These are outcomes other operators have published after digitising field-to-office documentation. We didn't run them, but they tell you what's on the table when the handoff gets cleaner. Groundline's job is to deliver the same mechanism with better tools and less friction.
- −80%
Dispatcher rewrite per job
10+ minutes down to under 2. Harris, mixed HVAC.
- −90%
Customer documentation requests
Cooper Oates, commercial HVAC.
- −65%
Invoice disputes
Design Mechanical, commercial HVAC.
- −93%
Invoice lag, completion to bill
~2 weeks to 1 day. Vari-Therm, residential HVAC.
- weeks → days
Billing cycle, step change
2.4-day cycle, +9% billed service time. Tampa Bay Trane.
Sources: Public case libraries: Harris, Cooper Oates, Design Mechanical, Vari-Therm, Tampa Bay Trane, ServiceTitan.
The bigger idea
HVAC is the beachhead. Service operations is the market.
Every service business, from plumbing and electrical to commercial facilities, fire and security, and industrial maintenance, runs on the same chain of handoffs. Intake. Dispatch. Execution. Documentation. Quoting. Invoicing. The pattern is identical; only the equipment changes.
Groundline is building the agent layer that rides on top of whatever system of record a company already has. Field signal in, structured output out, existing systems of record untouched. We are starting with closeout because it is the most expensive gap in the chain, and because once an agent can close out a job end to end, it can help run the next one.
The long view: multi-agent systems that handle the operational seams of a service business. Not a replacement for the software they already pay for. The connective tissue that makes it actually work.
Who we are
Applied AI that ships to production.
Engineering-led founding team out of UC Berkeley, Stanford, TUM (CDTM), and Cambridge. We have built the systems we are drawing on here.
- Contributor to DeepSeek
- Co-author of the StarCoder series
- Builder of CodeClash at Stanford
- Computer-vision and recommender-systems research
- Systems and process design at Amazon
- $10M+ measurable impact operationalising AI in reinsurance across location, cyber, and natural-catastrophe risk
Get in touch
Give us one branch. We'll give you the number.
Eight weeks. No fee. A verifiable ROI report at the end, whether you extend or not.