MEDICAL NEURO-SYMBOLIC ENGINE

MNSE The clinical AI that thinks like a specialist — and costs like a calculator.

Large language models are brilliant generalists that guess with probability. MNSE doesn't guess. It's a purpose-built neuro-symbolic engine for medicine — a small neural compiler fused with a deterministic symbolic resolver grounded in real drug, ICD-10, LOINC and SNOMED knowledge. The result: clinical-grade extraction, correction and now document generation that is auditable, hallucination-free in scope, runs on a single CPU, and costs $0 per call.

98.1% Certified on a frozen 10,000-case benchmark
$0 Cost per call — zero tokens, zero API spend
174,783 Indian drug brands in the live formulary
~120ms Full field extraction — on CPU, no GPU
99.98% Generative discharge summaries — 10,000 adversarial scans
A DIFFERENT SPECIES OF AI

This Is Not a Language Model.

For a narrow, high-stakes domain like medicine, a 175-billion-parameter generalist is the wrong tool. MNSE is the right one.

cloud A General-Purpose LLM

Trained on the whole internet to do everything, owned by no one

  • help_outlineProbabilistic — can hallucinate a drug that was never said and was never prescribed.
  • help_outlineBlack box — no trace of why a field was filled; hard to audit or certify.
  • help_outlineBilled per token, every single call — costs scale with every consultation forever.
  • help_outlineNeeds GPUs and a network round-trip; PHI leaves your building.
  • help_outlineNon-deterministic — the same input can return a different answer tomorrow.
  • help_outlineRetraining for your domain costs millions and weeks of GPU time.
VS
memory MNSE

A small neural compiler fused with a symbolic medical resolver

  • verifiedCannot recommend a drug that isn't in the formulary. Every output snaps to a known clinical concept.
  • verifiedGlass box — every result traces to a fired rule. Auditable by construction.
  • verified$0 per call, zero tokens. Tracked explicitly as a free local provider in the usage dashboard.
  • verifiedRuns on a single CPU thread — no GPU, no network, PHI stays inside the hospital.
  • verifiedDeterministic — the same input always returns the same, defensible answer.
  • verifiedTrains in ~11 minutes on one CPU from small, domain-specific data — not the whole internet.

MNSE is engineered to complement generative AI, not erase it. It owns the deterministic, repetitive, high-volume work (extraction, term-correction, drug safety) — and hands free-text prose generation to an LLM only where genuine language generation is required. Best of both worlds: the LLM writes; MNSE verifies.

HOW IT WORKS

Neural Where It Helps. Symbolic Where It Matters.

Messy speech goes in. Structured, formulary-grounded clinical facts come out — through three deterministic layers.

Layer 1 · Neural

The Concept Compiler

A tiny CPU-only encoder turns noisy transcripts and STT garble into discrete concept candidates with confidence scores — classifying every span as drug, lab, diagnosis, vital, frequency, schedule or other. It hits 99.91% validation accuracy across 88,810 examples, yet featurizes and runs in ~42 ms.

char-n-gram encoder 7-class softmax calibrated verifier no GPU required
south
Layer 2 · Symbolic

The Resolver Substrates

Concept candidates are snapped to canonical medical truth by a battery of deterministic substrates — bitmap algebra, lexicons and reference-range engines resolving in ~125 microseconds. This layer physically cannot emit a drug outside the formulary, a lab outside LOINC, or a diagnosis outside ICD-10. It either answers with proof, or returns nothing.

drug · 174,783 brands ICD-10 dx LOINC labs SNOMED · 90k+ concepts vitals · frequency · schedule drug-safety bitmaps
south
Layer 3 · Render

The Structured Output

Verified facts are assembled into 16 clinical fields — diagnosis, complaints, medications with typed dosing, investigations, follow-up and more — populated verbatim and traceably. Where free-text prose is needed, MNSE Generative now writes it: a maximum-a-posteriori template render over learned section models, with no sampling, so the same record always produces the same document. The doctor stays in control of the narrative.

16 structured fields typed drug entries MAP template render PII-safe fully auditable
CERTIFIED PERFORMANCE

Numbers, Not Vibes.

Every figure below is measured against frozen, version-locked benchmarks in the production repository.

Overall accuracy frozen 10,000-case drug + diagnosis benchmark 98.11%

9,811 of 10,000 cases passed at the latest deterministic-loop checkpoint.

Diagnosis recognition 3,000 dx cases 99.87%
Drug correction 7,000 adversarial brand cases 97.36%

Phonetic + perturbation + fuzzy matching across 174,783 indexed Indian brands.

Neural concept encoder validation set, n = 88,810 99.91%
AIDB adversarial defender suite locked attack corpus 100%

686 / 686 engineered adversarial drug attacks recovered.

Real-consultation field agreement 441 live consultations · structured parser 17.6% → 44.7%

Switching extraction to MNSE lifted overall field agreement 2.5× — diagnosis 37→76, drugs 2→64, examination 1→54.

Generative discharge summarisation 10,000 substrate-driven adversarials 99.98%

Corrupted, garbled and mis-sectioned scans of real discharge packets — 99.97% held on an unseen 3,000-case holdout.

NEW · MNSE GENERATIVE

It Writes the Document Too.

Extraction was only half the problem. MNSE now composes the clinical document itself — deterministically, with every sentence traceable to a source span.

description

Discharge summaries from a scan

A photographed or PDF'd discharge packet — skewed, garbled, sections out of order — becomes a clean, structured summary: diagnoses, procedures, investigations, dated medications with dose and frequency, condition at discharge and follow-up.

functions

MAP render, not sampling

Prose is produced by a maximum-a-posteriori template mixture over learned section models. There is no temperature and no sampling, so the same document renders the same way every time — a property no language model can offer.

fact_check

A verifier that knows when to stop

A Bayesian completeness verifier scores every rendered summary and either commits it or escalates. MNSE returning nothing is a designed outcome: it will not issue a document it cannot substantiate.

hub

Trained on mined terminology

SNOMED CT, ICD-10 and LOINC were mined into diagnosis, procedure, lab and substance substrates, then used to synthesise ten thousand realistic adversarial records — the corpus the engine was hardened against.

Every value in a generated summary is carried from a span in the source document. The engine has no capacity to invent a drug, a dose or a date that was not written down.

NEW · THE ENTERPRISE BRAIN

The Agent Runs on MNSE First.

Most of what an enterprise clinical assistant is asked to do is closed-world. MNSE answers that half outright, and hands the model only the questions that genuinely need one.

01

Intent routing

Twelve intents — open a view, launch a sub-agent, pull up a patient, generate a document, run a procedure guide — resolved deterministically from the doctor's words, with no model round trip.

MNSE · $0
02

Analytics narration

Counts, daily averages, busiest days, top diagnoses. The aggregation already holds the exact answer, so MNSE reports the numbers instead of paying a model to read them back — and cannot misquote them.

MNSE · $0
03

Document rendering

Discharge summaries, prescriptions, clinical notes and referrals composed from the record under a coverage gate, benchmarked line by line against a 3,280-patient reference corpus.

MNSE
04

Escalation

Open clinical reasoning, novel phrasing, genuine synthesis across free text — MNSE declines these by design and routes them to the language model. Knowing the boundary is the feature.

LLM
THE ECONOMICS OF DETERMINISM

Every Automation MNSE Owns Costs $0.

These are high-frequency operations that ran on a metered LLM. MNSE took them over — at a marginal cost of zero.

≈ 600×
cheaper per request than a GPU-served LLM — CPU symbolic resolution vs. metered token generation
spellcheckMedical term & drug correction
was: per-token LLM call
$0 / call

Now fully MNSE-authoritative in production. The certified 98.1% engine corrects every transcript before the doctor ever sees it.

dynamic_formTranscript field extraction
was: per-token LLM call
$0 / call

MNSE skips the LLM extraction call entirely and parses 16 structured fields itself — live on production today.

descriptionDischarge-PDF drug repair
was: per-token LLM call
$0 / call

Repairs every drug brand in OCR'd discharge summaries before and after prose generation — at no marginal cost.

savingsDrug-correction line item
was: ~⅓ of the AI bill
~33% bill removed

Drug correction alone accounted for roughly a third of the generative-AI spend. MNSE eliminated it.

Cost is tracked in the live usage dashboard, where MNSE is recorded as a $0 / zero-token provider distinct from metered LLM calls. The ~600× and ~33% figures reflect engineering measurements and deployment accounting; exact savings scale with your call volume.

THE SYMBOLIC BRAIN

A Substrate for Every Clinical Concept.

Each substrate is a specialised, deterministic resolver — composable, testable, and grounded in curated medical knowledge.

medication

Drug Resolver

174,783-brand formulary, phonetic + fuzzy matching, hallucination guardrail.

health_and_safety

Safety Substrate

Pregnancy, renal, hepatic & drug–drug interaction bitmaps from RxNorm / openFDA.

local_pharmacy

NLEM Overlay

NLEM-2022 essential medicines + India-specific contraindication rules.

coronavirus

Diagnosis Resolver

ICD-10 resolution with a fuzzy span scanner over a curated clinical seed.

science

Lab Substrate

LOINC analyte resolution, reference ranges, and HIGH / LOW / CRITICAL flags.

monitor_heart

Vitals Substrate

BP, pulse, SpO₂, temperature, weight, RR, RBS/FBS with clinical flagging.

schedule

Schedule Substrate

Natural-date parsing + OB/GYN procedure lexicon → appointment fields.

repeat

Frequency Substrate

BID / TID / OD / PRN / HS dosing, duration, route and meal relation.

account_tree

SNOMED Recognizer

90k+ concept hierarchy — disorder, finding, procedure & observable buckets.

blur_on

Embedding Fallback

Char-n-gram TF-IDF tier for misspellings & STT garble — under 1 ms.

inventory_2

Discharge Substrates

SNOMED, ICD-10 & LOINC mined into diagnosis, procedure, lab and substance indexes.

alt_route

Intent Router

Closed-world resolution of agent commands to twelve intents with typed slots.

edit_note

Narrative Composer

Turns dictation into third-person clinical prose — speech and speaker turns removed.

query_stats

Analytics Narrator

Reports aggregation results exactly; declines anything requiring inference.

LIVE IN PRODUCTION

From Spoken Word to Signed Record.

MNSE runs the deterministic spine of a real, full-scale clinical platform — every day.

mic

Capture

Doctor speaks naturally; transcript streams in.

spellcheck

Correct

MNSE repairs medical terms & drug brands at $0.

MNSE
dynamic_form

Extract

16 structured fields parsed — no LLM call.

MNSE
verified_user

Verify

Every drug snapped to the formulary & safety-checked.

MNSE
description

Generate

The discharge summary itself, written deterministically.

MNSE
assignment_turned_in

Finalise

Clean, auditable record — ready for the doctor to sign.

0 % Benchmark Accuracy
$0 Cost Per Call
0 + Symbolic Substrates
0 ms CPU Extraction

License the Engine. Own the Margin.

MNSE is available as a standalone, embeddable clinical-intelligence engine — drop it into any HIS, EMR or scribe and replace metered LLM calls with deterministic, auditable, $0 automations.

Message sent successfully!