MeetCarbon delivery method · AI forward deploy
General-purpose AI lowered the bar for “talking”
FDE sets the bar for “shipping”
FDE covers classic on-site system delivery. The core job is putting large models in the customer environment, wiring real data and permissions, and standing up business Agents so the front line actually owns usable AI — more than another public-web chat window.
- ScenesOffices · parks · stations
- AIPrivate models · Agent · RAG
- ConstraintsGov network · unified identity · data stays in-domain
- GoalInstall in the stack · use daily · own the capability
Deployment readiness · live loop
FDE sits inside the problems that actually stop the front line
Classic delivery makes the system usable. AI forward deploy also makes the model installable, the Agent wired, and the capability stay behind. Those issues rarely show in a demo; they surface on the customer intranet.
The model only chats in a public-web demo — how do you actually run inference on the customer intranet, government network, or a private stack?
01You wired a large model and it answers the wrong question or overreaches — how do permissions, data definitions and platform capabilities stay homologous with the conversation?
02Customers want energy-carbon weeklys, exception judgment and work assistance — how does the Agent sit inside the existing workbench?
03When the vendor leaves, prompts, knowledge bases and skill graphs leave too — how does the customer actually own operable AI?
04Three delivery modes — how close they sit to the problem
The industry is not short of model APIs or implementation manuals. It is short of engineering that installs AI in the customer’s own stack and grinds it until it works on site.
Pure product, remote
One capability, many customers
- HQ iterates from requirement docs
- AI stays in the public demo environment
- After go-live you queue on tickets
- A smooth demo ≠ front-line adoption
Classic implementation / consulting
Can configure and present; hard to change the product
- Click the config sheet and leave
- “Wired” is not the same as “usable”
- Agents struggle to embed in business systems
- Field learning rarely flows back to the platform
MeetCarbon FDE
One customer, many capabilities
- People sit in the real environment and intranet
- Deploy the model + build the Agent + wire permissions
- AI embeds in the workbench and daily workflows
- Capability can be handed over, operated and retained
Classic on-site system delivery still matters. FDE’s other core is “AI forward deploy” — putting large-model capability directly into the user’s internal systems, so users do not live forever on an external general-purpose assistant.
What MeetCarbon FDE is
Forward Deployed Engineer·Frontline deployed engineer
For MeetCarbon, FDE is both a role and a delivery promise: enter the customer’s real environment, make the business system daily work; and put large models, Agents and platform capabilities inside a boundary the customer controls, so AI becomes the organization’s own productivity — more than a one-off demo.
Software can be delivered remotely. Usable AI has to grow inside the customer’s systems.
People follow the project
Machine rooms, wards, duty rooms, the customer intranet — the workbench moves to where the problem happens.
AI integration also has to finish on the real network and permissions. A public sandbox that “looks chatty” is not enough.
Problems come from the front line
See how tickets are dispatched, how assessments are filed, how materials are submitted — then decide how the system and Agent grow.
The Agent skill list comes from front-line tasks, from the model vendor’s capability poster last.
AI installs into internal systems
Deploy inference, wire data and permissions, orchestrate Agents, embed in the workbench and business flows.
The goal is sustainable runtime in the customer environment — more than another external chat link.
Field learning feeds the product
Shared pains from offices, parks and campuses harden into MeetCarbon defaults and Atan skills.
The field is a gravel road; the platform is the paved highway. FDE turns the former into the latter.
FDE’s other core: put large models in the customer’s internal stack
Everyone can use a general-purpose large model. That does not mean the customer organization “owns” AI. MeetCarbon FDE finishes four jobs on site — deploy, use, build Agents, internalize capability — so Atan and MeetCarbon OS actually enter an environment the customer can control, operate and hold accountable.
AI does not replace statutory accounting, audit or engineering decisions. FDE turns “can talk” into “helps real work inside permission and data boundaries”.
How we deploy large models
Run inference inside a boundary the customer can accept
Choose public-cloud API, dedicated cloud or private inference from the customer’s security and network conditions. Finish gateway, GPU/compute, image and model-routing setup, intranet reachability and monitoring — do not hand keys to business users and say “call the API yourselves”.
- 01Assess: can data leave the domain, government/intranet isolation, latency and concurrency
- 02Land: inference / gateway / model routing into the customer environment or a dedicated channel
- 03Operate: health checks, rate limits and fallback, keys and audit trails
- 04Accept: real Q&A and timeout pressure tests inside the customer network
The field is the entry for product and Agent evolution
A good FDE model does not forever custom-build for one person. It hardens real constraints and high-frequency tasks found on site into platform defaults and Atan skills — so the next customer walks fewer detours.
Embed on site
Hear the business and network/security constraints; choose the model-deploy shape
Install in the stack
Inference reachable, permissions homologous, Agent in the workbench
Use in daily work
Train for adoption until the front line is willing to ask and to use
Feed the product
Shared skills enter the roadmap and reuse at scale
Four-step method: ship the system + install the AI
Hear the real need
Clarify business blockers and AI tasks together
The kickoff says “integrated control” and “add AI”. What actually shapes the product is the approval chain, assessment methodology, and what the front line asks and completes every day. System modules and Agent skills grow from there.
Get it right inside constraints
Design the system go-live path and the model-deploy path together
Unified identity, government network, masking and audit, security tests, and “can data leave the domain” all go in. First walk a path that can go live and an inference path that can be private or dedicated, then polish experience.
Close into one entry
Business portal and Atan share one org-permission set
Integration starts with one entry, one org, one accountability. Colleagues open one workbench to do facilities and energy-carbon work, and to ask Atan — answer permissions and page permissions stay homologous.
Stay after go-live
Follow system adoption and Agent ops together
The day the system runs and the model is wired is when real use begins. Whether rules need tuning, knowledge needs filling, which skills nobody uses — they become iterable versions, plus a capability handoff.
Go-live is the start. Delivery is when the front line actually uses the AI.
A week on site looks roughly like this
Days are not identical. A week mixes classic integration and process polish with live iteration on model routing, Agent skills and knowledge. The bar stays: smoother site, stabler acceptance, AI that actually helps.
Align blockers and AI tasks
Align business blockers and this week’s AI goals: which Q&A / work paths must run first, and what still blocks model deploy and permissions.
We hire people who close the loop on site
FDE stands between R&D, implementation and AI go-live. You do not need to do everything. What matters is willingness to be there, to talk, and to close models and Agents into the customer’s systems.
Accept residency, intranet integration, training and acceptance — work is not always at a desk.
Ask about business methodology, and whether data can leave the domain and who may see the answer.
From deploying the model to building the Agent, to front-line use, to a handoff the customer can run.
Field note: Wuchang smart facilities
Serving the Wuchang District Government Offices Affairs Center and advancing the “Wuchang District Smart Facilities Integrated Control Center”, engineers sat as close to the business as possible — first pulling fragmented, urgent, hard work into one entry. AI grew on the same org and permission system, without a second stack.
Repairs and property
Phone, WeChat and verbal handoffs mixed; dispatch was slow, progress invisible
Pulled into one ticket stack with SLA; staff can see progress
Meetings and visitors
Rooms booked solid and empty at once; the gate stalled between wait and security
Rules were ground into booking and access on site, not left on a blueprint
EV charging and energy-carbon
Order, fairness and abatement scoring stacked; materials were hard to file
One entry, data methodology that can roll up, filings that land when it counts
Government hard constraints
Unified identity, government network and security tests were “after go-live”
Constraints entered integration from day one; AI obeys the same identity and boundary
The above is project-progress language. It is not an official endorsement or administrative evaluation.
Atan and the OS are the capability surface; FDE is the delivery surface that installs them on site
Atan is the unified agent entry and persona. MeetCarbon OS / AI Runtime provides configurable models, Agent orchestration and homologous permissions. FDE installs all of it in the customer environment and grinds it until daily work.
Atan does not replace statutory accounting, audit or engineering decisions. FDE does not replace the customer’s business accountability. We deliver AI capability that can run, be operated and be handed over.
Make zero-carbon ship, and install AI on site
For institutions: you need engineering that can reside, deliver the system, and put large models and Agents in the internal environment. For talent: you want to make AI daily work at the customer site, more than closing a ticket.