“Why not just use ChatGPT?”
An LLM produces plausible commentary. SpecSense produces defensible engineering findings. That distinction matters when the deliverable is bound for site.
The one-liner
You can ChatGPT a document and get comments. You can't ChatGPT a comment register that survives a client audit. That's what we sell.
Where SpecSense earns its price.
Codified engineering judgment.
SpecSense encodes what a Chartered Engineer actually checks — the sequence, the priorities, the discipline-specific heuristics — not just what a general model happens to notice. Line class check before spec break check before PMS compliance. Line-list cross-check before flange rating verification. The order matters because rework cost compounds.
Traceable evidence, not opinion.
Every finding cites the code clause and the exact drawing location it comes from. ASME B31.3 §304.5.1, not “PWHT is usually required.” A discipline lead can accept or reject each finding on the evidence. An LLM produces confident-sounding paragraphs; a comment register survives contractual dispute.
Human-supervised escalation.
On supervised tiers, a Chartered Engineer triages severity, discards false positives, and signs the register. You are never shown machine output without human review. On indemnified enterprise contracts, the sign-off is bankable. An LLM chat log is not.
Discipline-specific workflows.
P&IDs, isometrics, datasheets, HAZOP registers — each has a canonical review flow, native format, and cross-reference set. SpecSense treats them accordingly. Uploading a P&ID and an equipment datasheet to a raw chat gives you commentary; uploading them to SpecSense gives you a cross-consistency check.
Accepted-finding feedback loop.
Every accepted or rejected finding tunes the severity calibration for future reviews. Over 100 reviews you build a per-client, per-discipline check library that is materially better than the day-zero baseline. A generic chat does not remember what you accepted last week; a workspace does.
Raw LLM vs. SpecSense — side-by-side.
| Capability | Raw LLM (ChatGPT / Claude / Gemini) | SpecSense |
|---|---|---|
| Format expected | Prose, paragraphs, occasional bullets. Verbosity varies. | Comment register in your template. Sortable by severity, discipline, drawing. |
| Evidence per finding | Rarely cites specific code clauses. Never cites drawing coordinates. | Code clause + drawing location + underlying calculation basis on every finding. |
| Severity calibration | Everything sounds equally important. Or equally uncertain. | Major / Minor / Info triage by a Chartered Engineer on supervised tiers. |
| False-positive handling | You have to manually re-check every claim. | Triaged before you see it. Rejected findings improve the check library. |
| Sign-off for contractual review | None. LLM outputs are not defensible in client audit. | Named Chartered Engineer signs the register. Indemnified sign-off on Enterprise. |
| Discipline coverage | Generic. Blends piping, process, mechanical vocabulary. | Discipline-specific check libraries: piping, process, mechanical, civil, structural. |
| Cross-document consistency | Struggles beyond a single document. Loses context across attachments. | Cross-checks P&IDs vs line lists vs equipment datasheets vs HAZOP registers. |
| Handling of native engineering formats | PDFs only. DWG, native P&ID files, HAZOP registers not handled reliably. | DWG, IFC, native P&ID exports, structured HAZOP registers, spec sections. |
| Data isolation & IP retention | Consumer-tier does not guarantee non-training. Enterprise tiers help but do not cover engineering-domain risk. | Per-tenant isolation, EU/UK/UAE residency, zero training on customer data, contractual IP retention. |
| Audit trail | Chat logs. Not structured, not exportable, not permanent. | Every access, review, and sign-off logged with actor and timestamp. Exportable audit trail. |
None of this is a criticism of LLMs. LLMs are our supply chain — we use them at the model layer and expect them to keep improving. The moat is everything wrapped around them.
“OK, but what about…”
“Can't you build this in a weekend with LangChain?”
You can build a demo in a weekend. You cannot build the codified check libraries, the discipline calibration, the accepted-finding feedback loop, or the Chartered-engineer sign-off in a weekend. Those are years of engineering practice compressed into a product. That is where the moat lives.
“LLMs get smarter every 3 months. Doesn't that eat your business?”
The opposite. Better base models make SpecSense stronger without changing the customer relationship. Model improvements are our supply chain, not our competition. See the roadmap page for the 5-year capability arc.
“We already have a discipline lead who reviews everything.”
SpecSense does not replace them. It catches the mechanical, high-volume checks so they spend their time on the judgment calls only they can make. Think capacity multiplier, not headcount replacement. The pilot proves the multiplier on your work in 48 hours.
“What if we just fine-tune a model on our SOPs?”
You will spend 6 months and end up with a specialised chatbot, not a comment register. The moat is not the model — it is the workflow, the sign-off, the audit posture, the EDMS wiring, the accepted-finding data. See our Buy-vs-Build breakdown on the roadmap page.
Prove it on one of your deliverables.
48-hour supervised pilot. Chartered-engineer sign-off. Cost credited to your first project pack.