OpsentinelOpsentinel
SpecGuardOpsMindInsight EngineAbout
Pre-Job · Document Intelligence

Catch the contradiction
before you quote it.

Read the RFQ before you sign it.

One missed tolerance contradiction in a $50K job is a rework loop that eats your margin for the year. SpecGuard reads RFQs and technical specs in about 30 seconds and surfaces every issue worth pricing around — or walking away from.

Start with 3 free analyses →Schedule a 30-min walkthrough

What is this?

SpecGuard reads RFQs, first-article inspection reports (FAIRs), engineering change orders (ECOs), non-conformance reports (NCRs), and material certificates — the multi-page PDFs that land in an estimator's or quality team's inbox every week. Median analysis runtime is 1.3 seconds in our internal benchmark suite (mocked LLM, 1 sample RFQ); add a 1–2 second LLM round-trip per chunk on real OpenRouter calls, so multi-chunk RFQs land well under a minute end-to-end. It outputs a list of specific things to flag before you commit to a price: contradictions between the print and the spec, undefined datums, tolerance stack-ups, missing surface finishes, ambiguous material grades.

Every finding cites the line or drawing region it came from. An estimator can verify any flag in seconds — so the conversation becomes “is this finding right?” instead of “did the tool make this up?”

Last benchmark: 2026-05-04 · 1.29s median across 5 runs of 1 sample RFQ. See apps/api/scripts/benchmark_specguard.py.

Built for

  • CNC and contract machine shops, 10–200 employees, $5M–$100M revenue
  • Estimators reviewing 5+ multi-page RFQs per week (drawings + spec packets + supplier requirements)
  • Aerospace / defense / oil-and-gas / medical work where one missed spec costs $20K+
  • AS9100, ISO 9001, IATF 16949, NADCAP-adjacent quality systems
  • Shops where senior estimators are the bottleneck and quote turnaround is a competitive lever

Probably not for you if

  • One-product OEMs running a single SKU at high volume
  • Shops getting one or two simple single-page RFQs a month
  • Teams without anyone reading PDFs end-to-end before pricing
  • Generic-AI tinkerers — SpecGuard is tuned for manufacturing specs, not a ChatGPT prompt

Walk through a real RFQ

1Drop in any RFQ PDF
opsentinel.io/specguard
Demo

SpecGuard · Upload an RFQ

6 risk categories ready

Drop your RFQ PDF here

Multi-page · drawings + spec packets · up to 50 MB

Multi-page PDFs, drawings, spec packets, supplier docs. No configuration.
2Six risk categories scan in parallel
opsentinel.io/specguard/analyzing
Demo

Analyzing · RFQ-2026-0142

14s elapsed · ~16s left

7 findings so far · 1 critical · 3 high

  • Tolerance contradictions
  • Datum reference closure
  • Surface finish completeness
  • Material grade ambiguity
  • Supplier flow-down clauses
  • Manufacturability checks
Contradictions, missing requirements, ambiguity, compliance, manufacturability, cost risk — all at once.
3A specific list of findings, severity-graded
opsentinel.io/specguard/results/rfq-2026-0142
Demo

RFQ-2026-0142 · 12 findings

30.4s · seed 0x4f2a · run #1881

1 critical4 high4 medium3 low
  • DTM

    Datum reference B not defined on any drawing surface

    Page 4 · Sheet 2 of 5

  • TOL

    Tolerance stack-up exceeds spec on bracket flange

    Page 2 · ±0.05 across two dims

  • FNS

    Surface finish callout missing on mating face

    Page 6 · figure 4

  • MAT

    Material grade ambiguous ("300-series stainless")

    Spec p.3 · 304 vs 316?

  • FLW

    Supplier flow-down for AS9102 not mentioned

    Cover page · clause 4.2

  • REV

    Revision letter on print and BOM disagree

    Print: Rev C · BOM: Rev B

Deduplicated, severity-ranked, every line cites a passage from the source.
4Click any finding for the citation
opsentinel.io/specguard/results/rfq-2026-0142#finding-1
Demo
Critical · DTMFinding 1 of 12

Datum reference B is referenced 3 times but never defined on any drawing surface.

Quoted from RFQ — Page 4, sheet 2 of 5

“Position tol .005 to datum [A][B][C].  Reference Datum B as established by mating surface — see sheet 2.”

Sheet 2 has no [B] callout on any face.

Operational impact

First-article rejection risk

Recommended action

Clarify with customer pre-quote

Determinism check: same finding on re-run · Two-pass adversarial verified

Verify in seconds. Same finding on re-run, two-pass adversarial verified.

A specific scenario

Wednesday morning.

An RFQ from a defense customer hits your inbox — 14 pages: drawing, spec list, supplier requirements, revision history. You'd normally block out 90 minutes to read it carefully before deciding to quote.

Instead you drop it into SpecGuard. Thirty seconds later you have 12 findings:

  • CRITICAL — Datum reference B is referenced 3 times but never defined on any drawing surface (page 4, sheet 2 of 5).
  • HIGH— Tolerance stack-up on the bracket flange is ±0.05 across two dimensions; the part can't be made to spec without a recalc.
  • HIGH — Surface finish callout missing on the mating face shown on page 6.
  • MEDIUM— Material grade specified as “300-series stainless” without specifying which — 304 and 316 are not interchangeable for the chemical-process environment described in the supplier requirements.
  • … and 8 more, each cited.

You decide whether to quote — and what to flag in your response — in 5 minutes instead of 2 hours. The finding list goes into your quote folder; the customer sees a more thoughtful response than they got from your competitors.

What you get

Six risk categories, every document

Contradictions, missing requirements, ambiguity, compliance gaps, manufacturability, cost-risk language — scanned per upload, no configuration.

Citations back to the source

Every finding quotes the exact passage in the PDF. The team argues about the finding, not whether the tool made it up.

Same document, same answer

Re-running an unchanged RFQ produces the same report. Determinism is auditable and defensible — by design, not by accident.

How it actually works

Document parsing

Multi-page PDFs, drawings with embedded text, spec tables, revision blocks. Text + structural-layout extraction in one pass. Pages that scanned poorly get explicitly flagged so the report can't silently skip a section.

Domain prompts

Manufacturing-tuned rules: GD&T resolution, datum-reference closure, tolerance stack-up sanity, surface-finish completeness, material-grade ambiguity, supplier-spec contradictions. Not “summarize this document” — specific pre-execution risks an estimator would catch on a careful read.

Citation enforcement

Every finding cites a passage from the source document. We check that each cited evidence string is grounded in the original PDF (≥70% token-overlap minimum, with a discriminative- token floor); findings the model can't ground get dropped during post-processing, never shown. Estimator verifies in seconds. Structured page/line/region metadata extraction is on the near-term roadmap; today's citations are free-text-with-grounding.

Severity rules

Critical / High / Medium / Low based on operational impact: will this stop the job, force rework, cost margin, or just need a clarification email? Mechanical issue types get deterministic severity (a numerical tolerance contradiction is always at least HIGH); soft issues like ambiguity cap at MEDIUM. Severity doesn't depend on the model's mood.

Determinism

Temperature-0 LLM, two-pass adversarial verification on every CRITICAL finding (refuted criticals downgrade to HIGH, never get silently dropped), canonicalization to dedupe near-identical findings. Same document → same report on re-run. Auditable, defendable, debuggable — and visible in the side-by-side diff when you re-upload a revised RFQ.

Under the hood

  • Document-type detection — automatically tunes the analysis based on whether you uploaded an RFQ, FAIR, ECO, NCR, or material cert
  • Auto-resolves PDF text quality issues — flags pages that scanned poorly so the report can't silently miss a section
  • Severity rules don't depend on the LLM's mood — mechanical issue types get deterministic severity
  • Two-pass adversarial verification on every CRITICAL finding (refuted criticals get downgraded to HIGH, never silently dropped)
  • Diff vs prior analysis — re-uploading a revised RFQ shows what changed, what got resolved, what got worse
  • Printable PDF report — operators print these for shop-floor meetings, attach them to customer emails
  • Forward an email to specguard@ — analysis lands in your dashboard within a minute or two (when configured)

What it doesn't do (yet)

  • CAD geometry analysis — we read drawings as documents, not as 3D models. On the post-pilot roadmap.
  • Revision-tracking across multiple RFQ versions — you upload v3; we don't know about v1 and v2 unless you upload them too.
  • Auto-pricing or quote generation — we surface risk; you set price.
  • Native ERP / quoting-tool integrations — drag-and-drop today; integrations on request.
  • Compliance-format export (AS9100, ITAR-controlled formats) — manual export only today.

See it on your own data.

Send a backlog of past RFQs — anything you analysed manually in the last quarter. We'll show you what SpecGuard would have caught, including the issues your team did catch (so you can calibrate it). Honest pilot, defended ROI number from your own work.

Start with 3 free analyses →Email us about a pilot

Pricing starts at $2,000/mo for SpecGuard alone (pilots from $99/mo for solo shops). See full pricing →

US-hosted · zero-data-retention LLM · NDA before any pilot · Security overview →