P1
Schema first
Every artifact is a typed record validated against a schema. Free text alone is never a deliverable.
// Agentic Engineering Framework · coming soon
AI × Engineering. Together, yet safe.
We're building AI agent processes that work inside your company's engineering processes, not around them. The responsible engineer gets AI support with clear guardrails: every output is a record that traces to a requirement, every deviation is explained, and every decision stays with them.
Agents assist. Engineers decide.
THE PROBLEM
Requirements, analyses and compliance evidence live in disconnected tools. Traceability is rebuilt by hand, defaults change silently and reviewers find gaps late. Unstructured AI use adds risk: fast output, weak provenance.
00 / YOUR PROCESS STAYS
An AI layer inside your process, not a replacement for it.
Your quality system, design reviews, configuration management and certification processes stay as they are. The agent processes plug into them so that engineers can use AI where it helps and still meet every obligation they already have.
STAYS WITH YOUR COMPANY
- Your engineering, quality and configuration-management procedures
- Design reviews, approvals and sign-off authority
- Safety classifications and certification decisions
- Your requirements and PLM tools as the master record
- Accountability of the responsible engineer
WHAT THE AGENTS ADD
- Drafts as structured, traceable records, ready for review
- Every deviation from a default explained before results
- Proposed trace links and gap reports for engineers to confirm
- Independent pre-review findings before your formal review
- A full audit trail of inputs, versions and model used
01 / PRINCIPLES
Six rules every agent works inside.
P2
Pipeline, not chat
Work moves through stages with entry and exit criteria. No agent skips a gate or approves its own output.
P3
Human authority
Named engineers approve baseline changes, safety classifications and certification decisions.
P4
Independence
Compliance Verification is read-only and shares no context or memory with the agents that author the work.
P5
Explain every deviation
Any departure from a default, baseline or standard method is flagged with its rationale, impact and affected requirements.
P6
Reproducibility
Every result can be regenerated from versioned inputs, code and tool configuration.
02 / PIPELINE
Mapped onto the lifecycle you already run.
The stages and gates follow your existing process. Agents prepare and check the work at each stage, and your existing reviews still decide whether it moves on.
- 01
Program definition
ConOps, mission profiles, scope, schedule and certification basis. Work is broken into board tasks.
Program Manager - 02
Requirements
One measurable "shall" per requirement, each with a verification method and a parent link.
Systems Engineering SME - 03
Validate in parallel
- Physics: feasibility & margins
- Airworthiness: clauses & MoC
- Safety: hazards & mitigations
- 04
Your baseline review
Your existing review and approval process signs off the requirement baseline. Agents can't pass it on their own.
Company approval - 05
Design tasks
Element-wise tools, analyses and reports, each run against exact baseline versions.
Design Engineer team - 06
Independent pre-review
An independent agent checks each deliverable before your formal review and raises findings, so reviewers start from a clean package.
Compliance Verification
OUTPUT Traceable evidence and compliance status, ready for your formal review and approval
03 / ROLES
An agent beside each engineering role.
Each agent supports a named human counterpart, holds only the context its task needs and has hard limits on what it may do. None of them signs anything.
Program Manager
Program context and coordination. The only role that creates and assigns tasks.
supports: Program manager
never writes requirements or approves deliverables
Systems Engineering SME
Derives requirements top-down, each with units, tolerance and a verification method.
supports: Chief systems engineer
never marks requirements approved
Design Physics Engineer
Checks requirements against first principles and shows the equation, inputs and margin.
supports: Discipline lead
never edits requirements directly
Airworthiness Expert
Maps certification clauses to requirements and means of compliance, and is only active when a certification basis exists.
supports: Certification lead
never cites clause text from memory as authoritative
Safety Engineer
Identifies failure conditions per flight phase and proposes severities and mitigations.
supports: Safety lead
never finalises a severity classification
Design Engineer
Runs one board task at a time in analysis or developer mode, with inspection-ready code.
supports: Design lead
never closes its own task
Compliance Verification
Independent, read-only review that raises findings with a severity and a required action.
supports: Independent checker / quality
never edits the work under review
04 / GUARDRAILS
Guardrails that raise confidence, not replace judgement.
The platform core is plain, qualifiable code: the data model, schema validation, gates, permissions and audit log. Agents connect through it and can't route around it. It connects read-only to the tools you already use, such as DOORS, Polarion, Excel and Git, and writes back only through your approved change process.
ANALYSIS MODE
Every parameter that differs from its default is reported before any result is shown.
DEVELOPER MODE
Readable code with explicit units, requirement IDs and a unit test per function.
- Schema validation on every write; invalid records are rejected with a reason
- Baseline control: changes merge only through an approved change request
- Role-based permissions; verification is read-only on design data
- Default-parameter watch blocks unexplained differences
- No gate closes with an unverified requirement, an unmitigated hazard or an open finding
- Only registered, versioned tools produce baseline outputs
- Human sign-off on baselines, safety, certification and final compliance
- Full audit log: who, what, when, inputs and model version
05 / DIGITAL THREAD
Every number traces back to a need.
- ConOps need
- Requirement
- Cert clause · Hazard
- Design parameter
- Analysis case
- Code version
- Result
- Report
- Verification finding
- Compliance status
Applicable standards follow the program's certification basis, e.g. ARP4754B · ARP4761A · DO-178C · DO-330 · MIL-HDBK-516.
06 / ROADMAP
Built in phases. Each has an exit criterion.
in_work
Foundation
Platform core, task board, Program Manager and Systems Engineering agents, and read-only adapters.
EXIT: a traced requirement set from a ConOps, or an in-progress baseline imported with a gap report
queued
Validation & safety
Design Physics, Safety and Airworthiness agents.
EXIT: requirements validated; hazards and clauses linked
queued
Design & verification
The Design Engineer team, independent Compliance Verification and report generation.
EXIT: one requirement set closed with verified evidence
queued
Certification pilot
A full digital thread on a real element, reviewed by an independent certification specialist.
EXIT: evidence package accepted without traceability rework
This doesn't replace your processes, your reviewers or your approvals. It gives the responsible engineer AI they can check, trust and sign off with confidence.