01
Prepare evidence
Start with a requirement, specification, test case, design record, or process artifact. Get a structured draft finding sheet with clause references and corrective-action suggestions.
Automotive software evidence · on-premise
Make the work assessor-ready.
Saakshya turns the engineering artifacts your team already creates into cited, reproducible draft evidence for ASPICE and the safety case. Your data stays on your hardware. A qualified human keeps the decision.
Offline by defaultDeterministic harnessNo compliance verdicts
SAakshya / review run 0042LOCAL
01Artifact receivedBMS requirement · content hash storedLOCKED
02Licensed corpus retrievedASPICE v4.0 · clause-level provenance3 HITS
03Draft analysis producedSchema-locked local model · temperature 0DRAFT
04Human review requiredFinding sheet · traceability · audit logPENDING
2bounded model calls per artifact
0cloud API calls by default
0.55retrieval abstention floor
1named human reviewer owns judgment
The product
A compliance harness, not an autonomous assessor.
The model is replaceable. The trustworthy system around it is the product: ingestion, retrieval, abstention, schema contracts, citation checks, persistence, and a trace a reviewer can replay.
02
Protect sensitive IP
Run the pipeline inside the customer network with a local open-weight model. Store a content hash and short excerpt—not the full source artifact—in the audit database.
03
Keep judgment human
Every result is stamped draft-only. The system refuses unsupported citations, abstains when context is weak, and cannot write “certified” or “compliant” into an output.
The harness
Fixed steps. Bounded cost. Reproducible evidence.
Saakshya is intentionally non-agentic. There is no model-controlled plan-and-act loop, no runaway tool calling, and no hidden third semantic attempt.
01
Normalize and hash. Reject empty or boilerplate artifacts; deduplicate without retaining the full IP-bearing text.
02
Retrieve and abstain. Fuse lexical and semantic search over the customer’s licensed corpus. If the evidence is weak, stop before calling the model.
03
Classify and analyze. Two schema-locked local calls: one process-area classification, one gap analysis. Zero gaps is a valid answer.
04
Validate and record. Cross-check clauses, reject malformed or banned output, append an audit event, and export a draft finding sheet.
Architecture decision
Harness first. Fine-tuning only if evidence earns it.
The current engine runs a base local model. Fine-tuning is not assumed to be the moat; it will be compared against strengthened RAG on an expert-verified gold set.
RAG + guardrailsQLoRA pilot
Standards are versioned behind an abstraction layer so an ASPICE v4.1 transition does not silently invalidate clause mappings. The latest research treats the 4.1 preview as a migration signal—not a reason to make unsupported claims today.
Assurance boundaries
Designed to support review—not replace it.
The latest requirements work narrows the claim: Saakshya reduces preparation time and improves traceability. It does not certify a process, qualify itself automatically, or become a vehicle safety function.
| Question a buyer will ask | Saakshya’s answer | Evidence supplied |
|---|---|---|
| Does it decide compliance? | No. It only produces draft evidence and advisory findings. | Immutable draft-only state, banned-verdict guardrail, human review record. |
| Does on-premise remove regulation? | No. Air-gapping protects IP; intended purpose still governs obligations. | Intended-purpose statement, model bill of materials, change and incident records. |
| Can it support ISO 26262 tool analysis? | It is designed to support a customer-specific TI/TD/TCL classification discussion—not to promise universal TCL1 acceptance. | Qualification-support dossier, failure modes, validation suite, reproducibility controls. |
| Does ISO/PAS 8800 govern this tool? | Not as an off-board engineering assistant; its scope is in-vehicle AI elements. | Clear product boundary and no vehicle-control function. |
Built for the evidence chain
From software artifact to reviewable finding.
The first wedge is BMS and ECU software. The architecture is intentionally useful wherever a regulated engineering team must connect work products to a changing standard.
SWE
ASPICE preparation
Start with SWE.1–SWE.6 plus selected SUP.1 and MAN.3 coverage. Bring the customer’s licensed standard and preserve clause-level provenance.
SAF
Safety and security bridge
Roadmap coverage for ISO 26262 Part 6 and ISO/SAE 21434, with separate evidence states and no automatic safety-case acceptance.
OTA
Continuous lifecycle readiness
Future FOTA/SUMS workflows can connect update evidence back to requirements, risk decisions, tests, and released software.
Product path
A measured route from preparation to lifecycle assurance.
Each phase has a narrower promise and a validation gate. The next phase starts only when the evidence and design-partner feedback justify it.
PHASE 01 · CURRENT
ASPICE workflow
Ingest artifacts, retrieve licensed clauses, classify process areas, draft gaps, suggest corrective actions, and preserve traceability.
PHASE 02
Compliance shield
Add ISO 26262 and ISO/SAE 21434 evidence workflows, tool-confidence support, and cross-domain traceability.
PHASE 03
FOTA readiness
Connect software updates to requirements, safety/security decisions, tests, release evidence, and UNECE R155/R156 readiness.
PHASE 04
Safe scaffolding
Generate bounded interface and test boilerplate—not safety-critical state-machine logic or unsupervised production code.
Straight answers
Questions we expect in a serious pilot.
Are you building your own language model?
No. The current slice uses a commodity open-weight model locally. The defensible work is the domain data, deterministic harness, evidence contracts, and verified evaluation—not pretending to own the weights.
Why not make it agentic?
For evidence preparation, reproducibility is more valuable than open-ended autonomy. A fixed sequence has known cost and enumerable failures. Agentic loops may be revisited only when a bounded, replayable pattern is justified.
Can Saakshya replace an ASPICE assessor?
No. It prepares and organizes evidence. An accredited or qualified assessor and the customer’s responsible engineers retain the judgment and sign-off.
Who should be the first design partner?
Tier-1 suppliers or OEM internal teams with BMS/ECU software, recurring assessment preparation, sensitive artifacts, and a domain expert who can create a small held-out gold set.
Design partnership
Bring one real artifact. Leave with a sharper evidence workflow.
We are looking for a small number of teams willing to show the real preparation pain, validate the draft output, and help us measure what actually improves.