Bayesian phase calc
Probabilistic CALPHAD
Phase fractions in steel + Ni alloys, with calibrated uncertainty bands. Engineers don't get a point estimate — they get a probability that their specific heat clears the spec.
How it works
- ·13 alloy systems calibrated to specific publications: Andersson 1987, Hammar 1979, Williams 1958, Pohl 2007, Hertzman 1985.
- ·Laplace approximation around the posterior mode + 4 000-sample posterior predictive draws for fast inference (~50 ms per call).
- ·Returns mean σ-phase / γ′ fraction + 95 % credible interval + decision verdict (P(σ ≤ threshold)).
- ·pycalphad TDB integration when available; empirical-equation fallback otherwise.
Why this matters: Thermo-Calc gives a deterministic phase fraction. We give the probability your production heat falls below your tolerance threshold — the number an engineer actually needs to sign a stamp.
Code-clause math
Pressure equipment
Direct closed-form ASME / API / EN checks for every pressure boundary. No FEA setup, no post-processing.
How it works
- ·ASME VIII Div 1 (UG-27 cylindrical, UG-32 ellipsoidal heads, external pressure) + Div 2 (nozzle reinforcement, fatigue Annex 5.B).
- ·B31.1 power piping · B31.3 process piping · B31.4 liquid pipeline · B31.8 gas transmission.
- ·API 650 new tanks · API 653 inspection / re-rate · API 661 air-cooled HEX · API 660 shell-and-tube.
- ·ASME III NB nuclear-class-1 stress checks; EN 13445 (PED), EN 13480 (industrial piping) European track.
Why this matters: ANSYS does FEA, you write the post-process yourself. Granta has lookup tables. We pre-package the actual code-clause math with the citation embedded in the verdict.
API 579 conversational
Fitness-for-service
Step-by-step API 579 walkthroughs that cite the specific clause at each branch. Real physics — not a yes/no decision tree.
How it works
- ·Parts 4 (general metal loss), 6 (pitting), 8 (weld misalignment), 11 (fire damage), 12 (dent / gouge), 13 (shell distortion), 14 (fatigue).
- ·Conversational session: each level (L1 screening → L2 engineering → L3 FEA) asks only for the inputs that level needs.
- ·Physics callbacks at each step — erosion-velocity, cavitation, thermal-shock — fire when the level requires them, with cited correlations.
- ·Verdicts cite the §clause; rationale is reproducible from frozen state.
Why this matters: Manual API 579 takes an integrity engineer a week per asset. Conversational walkthrough takes minutes. Same verdict, same citation chain, same defensibility.
API 581 RBI
Risk-based inspection
Probability of Failure × Consequence with full damage-factor breakdown — every contributing mechanism is listed and its DF rationale is auditable.
How it works
- ·PoF: thinning, SCC, brittle fracture, creep, corrosion-under-insulation, sour H2S, HTHA — each its own DF that combines into a total via API 581 §3.
- ·POD-adjusted DF reduces PoF based on the inspection effectiveness category (A best → E none) actually achieved on the last campaign.
- ·CoF: fluid hazard × inventory × hole-size scenario → area consequence (m²) + business consequence (USD).
- ·Combined risk → 5 × 5 matrix position → recommended next inspection date + NDE method, rendered as an inline matrix image.
Why this matters: Enterprise RBI software is offline-only and six-figures per seat. Same API 581 math here, with the risk matrix rendered next to the verdict.
The differentiator
Compliance bridge
MaterialState + ServiceConditions → cited code-verdict in milliseconds. This is the single thing no giant ships.
How it works
- ·Every clause is a small Python function: state + service → PASS / WARN / FAIL plus citation text.
- ·Joint probability P(PASS) propagates posterior uncertainty across every check — not just point estimates.
- ·Compliance diff between two MaterialStates (before / after MOC, code-edition delta).
- ·52 public-domain owner-spec overlays apply on top: NACE MR0175, API 941, ASME PCC-2, DNV-ST-F101, ISO 27913, EN 13445, FDA 21 CFR 211, …
Why this matters: Granta is a database — you read material values out. ANSYS is a solver — you compute stress. We're the verdict engine: state in, ASME-compliance out, with citation chain.
Auditable history
Time-machine replay
Reconstruct any historical verdict from its SHA-256 stamp — byte-for-byte, even years later. Litigation-grade.
How it works
- ·Stamp manifest captures: inputs hash + outputs hash + frozen code edition + frozen posterior version + MTC heat-genealogy chain.
- ·Replay rebuilds the verdict from the manifest using the same code that issued it. Reproducibility is mechanical, not approximate.
- ·RFC 3161 trusted timestamp from FreeTSA proves the moment-in-time on every stamp.
- ·Edition-delta tool diffs verdicts across code revisions (e.g. ASME VIII 2019 → 2023) for an entire fleet at once.
Why this matters: Best the giants offer is 'we'll search the email archive'. We give you a SHA-256 → exact verdict reproduction in seconds.
Tamper-evident
Audit hash chain
Every stamp is a hash-chained PDF. Modify a single byte and the chain breaks; an auditor can verify provenance without trusting our server.
How it works
- ·SHA-256 over canonical (sorted-key) JSON of every input, computed before the verdict is generated.
- ·SHA-256 over outputs + posterior trace + every embedded visual after generation.
- ·Replay snapshot includes all dependencies — TDB version, code edition, MTC OCR text, vendor-overlay revision.
- ·Stamp PDF embeds both hashes + RFC 3161 timestamp from a trusted authority, plus engineer-of-record license + jurisdiction.
Why this matters: A static PDF report has no provenance. Ours is a verifiable evidence package — third party can re-derive the verdict from the manifest, without contacting us.
Independent verification
Cross-vendor sanity
Every shipped output is benchmarked against an external reference. Disagreements are surfaced, not hidden.
How it works
- ·Lightweight emulators wrap common vendor outputs — Thermo-Calc TDB lookup, Aspen heat-balance, FactSage activity, AFT Fathom Δp.
- ·On every verdict: matseng prediction vs each emulator → quantified delta in % of measurement uncertainty.
- ·Per-mechanism breakdown when emulators disagree (e.g. our σ-phase posterior says 1.1 %, Thermo-Calc TDB says 0.6 % — here's why).
- ·Tolerance band configurable per parameter (default 20 %); outside-band disagreements get flagged YELLOW on the stamp.
Why this matters: A vendor that marks its own homework is a vendor you can\'t trust at audit time. We benchmark every output against an external reference — and tell you when we disagree.
Cited-clause RAG
Standards retrieval
403 cited code-clause chunks indexed in a TF-IDF + semantic search corpus. The agent quotes the actual clause text and cites the §number every time it makes a verdict.
How it works
- ·STANDARDS_KB ships 403 chunks: ASME VIII Div 1+2, B31.1/3/4/8, ASME IX, API 510/570/581/653/941/945, NACE MR0175/MR0103/SP0472, BS 7910, DNV-ST-F101, ISO 27913, EN 13445/13480.
- ·Each chunk is structured (standard, clause, title, text, formula, edition, effective_year) — not blind text-blobs.
- ·TF-IDF retrieval over the chunk corpus + semantic re-ranking surfaces the right clause for the question.
- ·No cloud-side embedding API needed; the index is part of the platform.
Why this matters: Most agents hallucinate code-clause text. Ours quotes the chunk and shows the §number. The verdict is grounded in the same clause an auditor would open the codebook to.
BYO-API
Integrated AI agent
264 callable tools, one chat interface. Your API key, never ours. Engineering Stack-Overflow with the platform's full toolbox underneath.
How it works
- ·TF-IDF + synonym-expansion routing: 264 tools indexed, agent sees the top 12 most-relevant per prompt (~3 K tokens instead of ~100 K).
- ·Anthropic prompt caching on system + tool block — 0.1× pricing on cache hits, drops typical-prompt cost from $0.30 to $0.03.
- ·Direct browser-to-Anthropic SSE streaming. Your key + your prompt never touch our backend.
- ·Agent loop executes tools end-to-end (RBI → matrix-visualize → stamped package → replay manifest in one prompt) with self-healing retry on routing miss.
Why this matters: Giants still ship 1990s GUI workflows. The agent turns 264 specialized engineering tools into one usable interface — the platform is an API + a chat, not a desktop install.