MEDHA is DataSwitch's private Small Language Model (SLM) designed for enterprise data engineering. It runs CPU-only with ~50ms latency, zero token cost, and can be deployed fully air-gapped with no data egress.

What is the difference between MEDHA and Trident?

MEDHA is a small, fast model for high-volume bounded work like vocabulary matching, classification and routing at CPU speed. Trident is a larger model for complex adjudication and repair, invoked only when deeper reasoning is required.

Can MEDHA and Trident be deployed air-gapped?

Yes. MEDHA runs entirely on-prem or in private cloud with zero data egress. Trident can operate with error-only exchange, ensuring source logic is never exposed externally.

Platform / Private AI

MEDHA & Trident. Private AI That Never Leaves Your Walls.

Deterministic small language models built for enterprise data engineering — running CPU-only, air-gapped, at zero token cost.

~50ms latency0% hallucinationno GPUno data egressno per-token bill

~50ms

Inference latency (CPU-only)

Hallucination (deterministic output)

Per-token cost

Air-gap

Deployable (zero data egress)

Two Models, One Engine

A Small Model for Speed. A Large One for Depth.

MEDHA handles high-volume, bounded work at CPU speed. Trident takes on the heavy reasoning. Together they cover the full data-engineering workload without sending a byte to an external service.

MEDHA

Small Language Model

The fast, deterministic router

Runtime: CPU-only, no GPU dependency

Latency: ~50ms per inference

Cost: Zero token cost

Deploy: Air-gapped inside your environment

Work: Vocabulary, matching, classification, routing

Trident

Large Language Model

The heavy-reasoning specialist

Runtime: GPU or API, error-only exchange

Role: Complex adjudication & repair

Privacy: Source logic not exposed externally

Trigger: Invoked only where depth is required

Output: Validated, deterministic result

Technical Specification

Built for Regulated Estates.

Property

MEDHA

Model class

Private Small Language Model (SLM)

Execution

CPU-optimized — no GPU required

Latency

~50ms per inference

Determinism

0% hallucination, 100% syntax-validated

Deployment

On-prem · private cloud · fully air-gapped

Data egress

None — schemas & code stay in your environment

Token cost

Zero — no per-token billing

Governance

MCP & GraphRAG over an RDFS/OWL knowledge graph

Why Private & Deterministic Wins

The Opposite of a Probabilistic Cloud LLM.

sovereignty

Data never leaves

Air-gapped execution means proprietary schemas, code and lineage stay inside the enterprise. Built for BFSI, healthcare, aerospace and public-sector estates.

economics

No GPU, no token bill

CPU-only inference removes GPU capital cost, and zero per-token pricing makes high-volume conversion — and SME-scale economics — actually viable.

trust

Deterministic, not a guess

Rule-governed output validated against the target grammar. 0% hallucination isn't a benchmark — it's a property of how the models compile.

The engine that runs on them

SwitchIE.

MEDHA and Trident power SwitchIE, DataSwitch's agentic data engineering engine — turning legacy estates into validated, platform-native code across 15–30 tools.

Run private AI on your own hardware.

See MEDHA convert a representative workload air-gapped — deterministic output, zero egress, measured on your estate.