The FDE method
FDE stands for forward-deployed engineer: our engineers work at your site, build AI into everyday work alongside your people, and are accepted against agreed metrics.
Four stages, each with a gate
1
Align on the use case
Use-case list, success metrics, data and system boundaries
Gate: No metric written down, no PoC
2
Build and launch together
Working system, connectors, evaluation set, acceptance report
Gate: Below the bar, no launch
3
Run in production
Monthly results report, knowledge operations, model tuning
Gate: Metrics slip, go back a stage and fix
4
Reuse and expand
Connectors, industry templates and evaluation sets reused
Gate: Similar use cases get faster each time
Six places where AI projects stall
| Hurdle | What it looks like | What we do |
|---|---|---|
| No defined use case | It starts with “we need AI” — nobody can say which step it saves or who owns it | A use-case workshop: write the success metric first, then build |
| Data in silos | Knowledge is scattered across drawing vaults, file shares, laptops and chat groups | Connect native systems on site, index through file encryption, and keep each person's permissions |
| Legacy systems without APIs | ERP, MES and CAD offer no API — only spreadsheet exports and manual entry | Connectors plus deterministic write-back, without modifying your core systems |
| Wrong acceptance yardstick | Only speed gets measured, not accuracy — one wrong answer and nobody trusts it | Build the evaluation set from your real material and put accuracy into the acceptance bar |
| The business doesn't adopt it | The pilot passes in IT, but no one in the business uses it | Ship it inside WeCom, Feishu, DingTalk and existing software — no install, no training |
| Unclear boundaries | The language model is allowed to produce numbers, issue commands or drive equipment | The model only proposes; parameters, checks and database writes are decided by code and human approval |
The model only proposes
The language model understands and explains; numbers, parameters and writes are decided by code and people.
1
An employee asks
in WeCom, Feishu, DingTalk or business software
2
Language model
understands, retrieves, drafts — proposes
3
Rules engine
computes numbers, checks parameters, traceable
4
Human approval
owners confirm high-risk actions; submitter ≠ approver
5
Write back
to ERP, drawing vault, task system — fully audited
Safety and control
- Data stays inside
- Privately deployed open-weight Chinese LLMs; external APIs can be disabled; the model gateway is the only exit and logs everything.
- The model only proposes
- Understanding, retrieval and drafting go to the model; calculations, parameter checks and writes to business systems are decided by deterministic code and human approval.
- Permissions and separation of duties
- Permissions follow the source system, filtered per person; high-risk actions need approval, and the submitter can't approve their own request.
- Full audit trail
- Every question, call and change is recorded; key figures trace back to the original documents.
- Domestic-stack ready
- Runs on domestic databases and ARM servers, with offline installers that set up from scratch without internet.
Deployment options
- On-premises
- In your own data center or servers; data never leaves the company.
- Dedicated cloud
- On the dedicated or government cloud you choose, scaling as needed.
- Hybrid
- Sensitive data and models stay local; non-sensitive workloads use cloud resources.
Where people use it
- WeCom
- Feishu
- DingTalk
- Web
- Mini programs
- Embedded in existing software
How we work together
1
Validate the use case
A PoC on your real material, with acceptance metrics agreed up front
Working PoC, evaluation report, launch plan
2
Deliver
Connect systems, go to production, train the business users
Production system, connectors, evaluation set, acceptance report
3
Operate
Knowledge operations, results monitoring, model tuning, new use cases
Monthly results report, ongoing improvement, priority support
- Nothing launches below the bar; results are reviewed monthly after launch
- On site for alignment and launch; remote operation after that
- Each use case is quoted on its own — never billed by the person-day
Questions
Will our data leave the company?
By default everything is deployed privately — models and data stay on your own servers, and external model APIs can be disabled in one step. A dedicated or government cloud of your choice also works.
Do we need to buy GPU servers first?
Not necessarily. The PoC can run in our demo environment on de-identified material; once results are confirmed, we size compute to the use case, starting from consumer-grade GPUs.
How soon will we see results?
A half-day workshop, a PoC proposal within a week, the PoC in 2–4 weeks; after acceptance, production in 4–6 weeks.
How many of our people are needed?
One business owner and one IT contact. During the PoC you provide real material or data samples and take part in reviews.
What if the model gets it wrong?
Key figures are computed by code, and the model only understands and explains. Answers carry their sources so they can be checked, and high-risk actions always need human approval.
Can it connect to our existing systems?
Yes. Connectors reach ERP, OA, drawing vaults, file shares and business databases — read-only where possible, writes only through deterministic rules and approval.
Can it run on a domestic stack, offline?
Yes. It runs on domestic databases and ARM servers, with offline installers; install, upgrade and rollback have all been rehearsed.
How is it priced?
Per use case: the PoC as a project, delivery as implementation plus licence, operations as an annual subscription. We don't bill by the person-day.
Can you do something that isn't on the list?
Yes. The list covers common use cases; our delivery is tailored to each client anyway. Start with a half-day workshop to pin down the need and the acceptance metrics.
Start with a half-day workshop
Your business owner and IT lead join us to pin down 1–3 candidate use cases and their success metrics.
Book a use-case workshop1
Use-case workshop
- Business owner and IT lead attend
- Map 1–3 candidate use cases and the state of the data
- Output: use-case list and draft success metrics
2
PoC proposal
- Data and system boundaries, evaluation-set design
- Milestones, commitments on both sides, acceptance bar
- Output: a PoC proposal ready to sign
3
Run the PoC
- Your real material, in a private or demo environment
- Weekly check-ins with visible metrics
- Output: evaluation report and launch plan