The Math Copilot for AI Agents

Your LLM groks. Grokster verifies.

LLMs think. Pramana proves. Add deterministic constraint verification to any AI agent — sub-millisecond, zero hallucination risk, works alongside your existing stack.

Try Live Demo View on GitHub
<1ms
Verification Time
100%
Deterministic
0
LLM Dependencies
Benchmark Tests
The "Aha" Moment

Your LLM understands intent. Pramana catches the math.

LLMs are great at reasoning about context. But numeric constraints need a different kind of check.

LLM Alone: "Looks reasonable"

// Agent generates patient dosage { "patient_weight_kg": 70, "dosage_mg": 850, "max_safe_mg_per_kg": 10, "heart_rate": 142, "hr_limit": 120 }
PASSED (LLM: "values look reasonable")
The LLM understood the context perfectly — but it doesn't do math. Dosage is 12.1 mg/kg (21% over limit). Heart rate 18% over threshold. Two numeric violations invisible to language models.

LLM + Pramana: CAUGHT

{ "verdict": "FLAGGED", "confidence": 0.19, "violations": [ "dosage 850 exceeds limit 700 (12.1 mg/kg > 10)", "heart_rate 142 exceeds limit 120 (+18.3%)", "cross-pressure: 2/2 dimensions critical" ], "time": "0.3ms" }
CAUGHT — 2 violations, 0.3ms, deterministic
Your LLM handles intent and context. Pramana handles the numbers. 850 > 700 is always true — every time, on every run, in under a millisecond.
The Gap

LLMs are brilliant. But they can't do math.

Language models excel at reasoning, intent, and context. But constraint verification needs deterministic math, not probabilistic language.

🎲

Fuzzy on numbers

LLMs approximate. "850 mg looks reasonable for a 70 kg patient" — but 850 > 700 is a math problem, not a language problem.

🐢

Expensive for simple checks

Using a $0.01 LLM call to verify "is X > Y?" is like hiring a lawyer to check your arithmetic. Use the right tool for the job.

🤝

Better together

Your LLM owns intent, context, and nuance. Pramana owns boundaries, constraints, and numeric proof. Together: production-ready.

The Pramana Pipeline

LLM generates. Pramana verifies. Action executes.

A lightweight verification layer that sits downstream of your LLM — three deterministic gates powered by mathematical constraint checking.

1

Evidence Gate

Nyaya

Classifies evidence by type (direct, inference, analogy, testimony), detects logical fallacies, computes weighted reliability.

2

Constraint Gate

Vedic Sutras

16 mathematical verification methods — boundary checks, equilibrium, proportionality, completeness, gap analysis.

3

Strategy Gate

Arthashastra

Risk assessment using strategic postures, resource evaluation, and cost-benefit analysis. Context-aware action gating.

How Nyaya Logic Maps to the Three Gates

Nyaya (Sanskrit: nyāya, "method of reasoning") is a 2,500-year-old system of formal logic that classifies knowledge claims by their pramana (means of valid knowledge): pratyaksha (direct observation), anumana (inference), upamana (analogy), and shabda (testimony). Gate 1 applies these classifications to weight evidence reliability. Gate 2 uses Vedic mathematical sutras as constraint-satisfaction patterns — each sutra maps to a verification method (boundary check, equilibrium, proportionality). Gate 3 draws from Kautilya's Arthashastra for strategic risk assessment: shadgunya (six postures) and shakti traya (three powers) determine whether an action's risk profile warrants execution. This isn't decoration — it's the actual logic the engine runs.

The 16 Vedic Sutras as Constraint Checks

Each sutra maps to a verification method that checks a different property of your data. The engine auto-detects which sutras apply based on your data shape, then computes an "energy score" (lower = better).

Nikhilam
Complement check — how far from the limit?
Urdhva Tiryak
Multi-dimensional — all constraints simultaneously
Shunyam
Equilibrium — do inputs balance outputs?
Anurupyena
Proportionality — do ratios hold across scale?
Sankalana
Additive — do parts sum to the whole?
Puranapurana
Completeness — are all required fields present?
Chalana
Delta analysis — are changes within bounds?
Yavadunam
Gap analysis — measure and bound the deficit
Ekadhikena
Sequential — each step must build on the last
Paravartya
Transpose — transform to simpler equivalent form
Vyashti
Compositional — parts consistent with the whole
Sheshanyankena
Residual — modular arithmetic verification
Sopaantya
Boundary — edge cases remain valid
Ekanyunena
Decremental — diminishing returns check
Gunitasamuchya
Distributive — aggregation preserves truth
Gunaka
Factorization — decomposition is valid

Three lines to verify anything

Add Pramana downstream of your LLM in three lines. Works with LangChain, CrewAI, AutoGen, or raw API calls.

verify.py
from pramana import Pramana

engine = Pramana()

# Verify constraints before acting
result = engine.verify_quick(
    data={"cpu": 0.82, "memory": 0.71},
    constraints={"cpu": 0.85, "memory": 0.90}
)

if result.verdict == "VERIFIED":
    execute_action()
else:
    halt(result.reasoning)

# LangChain integration:
from pramana.integrations.langchain import PramanaGuard
chain = my_llm | PramanaGuard(
    banned_patterns=["password", "ssn"]
)

Try it live

Paste JSON data and constraints. Hit Verify. See the result in real-time from our API.

Click "Verify" or try a preset...
Why Both

The right tool for each layer of safety

LLMs and math-based verification each have strengths. The question isn't which one — it's how to combine them.

⚠ LLM Alone

Great at intent & context
Fuzzy on numeric constraints
200-2000ms per check
$0.01-0.05 per verification
Can hallucinate safety verdicts
Non-deterministic output

⚠ Math Alone

~ 100% deterministic
~ Sub-millisecond
~ No semantic understanding
~ Can't parse intent or tone
~ Rigid without context
~ Handles structure, not meaning

✔ LLM + Pramana

LLM handles intent & nuance
Pramana handles constraints & math
Sub-ms verification layer added
$0.0001 per constraint check
Full audit trail on every decision
Deterministic where it matters
Built for production AI systems
128
Benchmark tests passing
100%
Recall on benchmark suite
<1ms
Average verification time
"GPT-4 handles our reasoning and intent parsing beautifully. But it was missing 40% of numeric constraint violations. Adding Pramana downstream took 20 minutes — now we catch 100%, in under a millisecond." — Barbarian Labs, multi-agent production system (content generation + autonomous pipelines)
"We didn't replace our LLM guardrails — we added Pramana alongside them. The LLM catches semantic issues, Pramana catches the math. Two near-misses on dosage limits would have been caught from day one." — Early adopter, healthcare AI safety team
Industry Solutions

Tuned for your domain

Pre-built verification profiles with industry-specific thresholds. Deploy in minutes, not months.

Pramana Guard

AI Agents & LLM Safety

Deterministic constraint layer for every AI action. Your LLM reasons, Pramana enforces the boundaries.

  • LLM output format & content verification
  • Agent action risk assessment
  • Prompt injection detection
  • Autonomous workflow gating
Pramana Comply

Financial Services

Auditable verification for regulated environments. Every decision traced back to evidence and rules.

  • Transaction anomaly detection
  • Balance reconciliation verification
  • AML/KYC evidence gating
  • Trade compliance checks
Pramana Clinical

Healthcare & Life Sciences

Zero-tolerance patient safety verification. Deterministic checks that regulators can audit line by line.

  • Vital signs boundary monitoring
  • Dosage constraint verification
  • Clinical evidence evaluation
  • FDA-auditable decision trails
Pramana Edge

Autonomous Systems & Robotics

Sub-millisecond pre-action verification for systems where latency kills. Runs on-device, no cloud required.

  • Pre-action safety verification
  • Sensor data constraint checking
  • Multi-dimensional boundary monitoring
  • Edge deployment — no network needed
Pramana Code

Construction & AEC

Automated building code compliance. IBC, OSHA, and ADA checks encoded as verifiable constraints.

  • IBC occupancy & fire safety verification
  • ADA accessibility compliance
  • OSHA safety constraint checking
  • Permit application pre-validation
Pramana SOC

Cybersecurity

Evidence-based incident response gating. Classify, verify, and gate security actions with forensic-grade provenance.

  • SIEM alert evidence evaluation
  • Automated response action gating
  • Incident evidence classification
  • Zero-trust verification chains
Pricing

Start free. Scale without surprises.

No per-token pricing. No surprise bills. Predictable costs at any scale.

Core

Open source, self-hosted

$0
forever
  • Full Python SDK
  • All 3 verification gates
  • 8 domain profiles
  • LangChain integration
  • CLI tool included
  • Community support
View on GitHub

Fortress

Enterprise & regulated industries

From $1,500/mo
unlimited verifications
  • Everything in Shield
  • Dedicated instance
  • Custom domain constraints
  • SOC 2 audit logs
  • On-premise deployment
  • Dedicated support engineer
Contact Sales

Our LLMs were brilliant — but they couldn't do math

We run a multi-agent AI system in production — autonomous agents handling content generation, data pipelines, and external communications. Our LLMs were great at reasoning about what to do. But when it came to numeric constraints — dosage limits, budget thresholds, resource caps — they got it wrong 40% of the time.

So we built Pramana as a complementary layer — grounded in mathematical constraint verification. The LLM still owns intent and context. Pramana owns the numbers. Together: 100% recall on constraint violations, sub-millisecond, and our LLMs are free to do what they're actually good at.

Now we're open-sourcing it so every team shipping AI agents can add a math co-pilot that never hallucinates.

Barbarian Labs — Minneapolis, MN

Explore what Pramana can do for your business

We work with teams to build custom verification profiles for their domain — from healthcare dosage checks to financial compliance rules. If you have a use case where AI decisions need to be provably correct, let's talk.

Start a Conversation See Industry Solutions
Get Started

grok + grokster = Verified.

Add your math copilot in under 5 minutes.

pip install pramana-engine Talk to Us