Forward Deployed
Engineering
that no search engine will ever reach.
Even the most capable model is only the drilling rig. The gold is extracted on site — where confidential design assets and the tacit knowledge of field engineers lie buried — by domain specialists led by an AI Champion.
Even the most advanced AI needs a partner in automotive
The AI industry has already converged on one answer: go direct where outputs can be verified, and partner with domain specialists where the gold lies unverifiable. Automotive is that gold.
Where it can be verified, they go direct
Coding, law, biology — domains where inputs and outputs are unambiguously verifiable. Builders of powerful AI enter these directly, without partners, because the model can check its own answers and learn without limit.
RLVR · Verifiable RewardBut the real value sits below the surface
The knowledge that makes verification possible exists only in public data. What actually runs an industry lives in the tacit knowledge inside expert minds and in client confidential assets — gold that no search will surface.
Tacit KnowledgeIn automotive, almost everything is confidential
Design documents, requirements and control logic are almost entirely proprietary, beyond the reach of any crawler. The more powerful the model in hand, the more urgently one needs a domain specialist who knows where the gold is buried.
Domain GapThe drilling rig (the foundation model) already exists.
The gold is extracted by the specialists who go on site.
— Those specialists are PopcornSAR.
The rig exists. The mining is ours.
Holding a foundation model and knowing the automotive domain are two different capabilities. PopcornSAR holds both — eleven years of embedded and AUTOSAR engineering, and PARVIS, our own AI toolchain.
Foundation Model
World-class foundation models. Powerful — yet blind to the automotive domain and to the confidential floor where the work happens.
the rig
A domain team led by an AI Champion
Eleven years of automotive embedded and AUTOSAR engineering, plus PARVIS. An AI Champion fluent in both domain and AI leads on site, holding the verification tooling and extracting the gold.
on site
The client floor
Confidential design assets and the tacit knowledge of field engineers. Real value emerges only here, where nothing is searchable.
Every domain has a grader.
Every domain but automotive — until now.
One criterion decides which domains AI enters directly: whether correctness can be graded mechanically. PopcornSAR exists to install that grader in automotive.
How Forward Deployed Engineering works
Nothing leaves the client boundary. Every AI artifact must pass a human verification gate.
On site, led by an AI Champion
An AI Champion fluent in both domain and AI embeds inside the client network and leads the team. One person owns the work from problem definition through verification, and confidential material never leaves the room.
Deployment without exposure
Three on-premise configurations matched to the client environment — private cloud, lightweight Mac Studio, and NVIDIA server room. Local LLMs supported.
The verification gate
AI-generated artifacts must clear deterministic tool validation and expert-plus-client review (HITL). Safety and real-time code always requires final expert approval.
Tacit made explicit
Scattered design assets and field knowledge are consolidated into Automotive SPICE-conformant documentation — including reverse engineering back up the V-cycle.
On these unsearchable floors,
PopcornSAR is already forward deployed
Recovering requirement specifications and test cases from production control source code
- Task
- Production BMS code with upstream documentation lost; the rationale survived only in one engineer's head
- Team
- Source sample received under NDA; on-premise analysis pipeline stood up
- Result
- Draft requirement specification and test cases within one week; technical validation passed
Automating MATLAB/Simulink ↔ C code conversion and diagnosis
- Task
- An organisation mid-transition to MBD; models and legacy C coexist, with consistency checked by hand
- Team
- Simulink and Stateflow environments built in-house; conversion and diagnosis workflow designed
- Result
- Model-to-code cross-diagnosis automated; migration risk surfaced early
Translating natural-language intent into CAPL test scenarios
- Task
- A verification team dependent on CANoe and CAPL, with scripting bottlenecked on a few experts
- Team
- Intent-to-CAPL translation pipeline with a measurement cross-check gate
- Result
- Non-specialists can now express scenarios; experts concentrate on review
Generating ARXML from natural language, validated against schema and custom rules
- Task
- ARXML conventions differing by release and internal rule — undocumented tacit knowledge
- Team
- NL-to-architecture and SW-component generation, double-validated by schema and custom rules
- Result
- Multi-release design automation; artifacts passing interface validation
Generating supplementary domain data as co-input, verified through client HITL
- Task
- Source requirements alone were insufficient as AI input; domain knowledge scattered across teams
- Team
- Domain engineers hand-authored supplementary data as co-input, with a client review loop
- Result
- Verification automation established under air-gapped conditions; client approval gate institutionalised
Auto-generating an SAD and 29 SDD documents while preserving full traceability
- Task
- A full English design document set for an overseas client — unachievable by hand within schedule and consistency limits
- Team
- SAD + 29-SDD generation pipeline with a traceability validator and anti-hallucination gate
- Result
- Full English artifact set generated; 100% traceability validation passed
Not generic FDE. Automotive FDE.
Few companies can transplant the verifiability AI depends on into a domain where the documents themselves are confidential.
Verifiability, transplanted
Tool-in-the-Loop deterministic gates make quality reproducible in automotive.
Confidentiality preserved
On-premise and local LLM deployment keeps data inside the client boundary.
Standards-conformant
Delivered on processes aligned to Automotive SPICE, ISO 26262 / 21434 and ASIL.
Led by an AI Champion
People fluent in both domain and AI lead the floor. Tools do not own outcomes; people do.
Domain & language coverage
Eleven years in automotive, operating across Korean, Japanese, Chinese and English.