Korvo Intelligence Engine

Meet Medha.

मेधा - Sanskrit for intelligence

On-device metrics for measuring retrieval and evaluation workflows - deterministic, token-free, and completely offline.

Models generate. Medha measures. You decide.

Or download Korvo free →

No account needed — the download starts right away.

Bundled with Korvo · Human-labelled release evaluation in progress

The Problem

AI models generate confidently.
Nobody checks their work.

You send a question to GPT-4, Claude, and Gemini. You get three plausible, confident answers. Which claims are actually agreed upon? Where do they contradict? What's uncertain? Today, you either read all three manually - or send them to yet another cloud model for synthesis, paying more tokens and waiting more seconds.

Cloud synthesis is slow and expensive

Sending multiple AI drafts to another LLM for evaluation costs tokens, takes 5-10 seconds per call, and requires internet. At scale, it's unusable.

No structured disagreement analysis

Models give you text. They don't give you "Claim A from GPT-4 contradicts Claim B from Claude with 73% similarity." You're left reading and comparing manually.

Evaluation at scale is a cost wall

Running 100+ experiment iterations overnight? At $0.03-0.10 per evaluation, that's $3-10 per run. Most people just... don't run experiments.

Architecture

Where Medha sits in your workflow.

Medha doesn't replace your AI models or human review. It provides deterministic, local measurements that make retrieval experiments reproducible.

Your AI Models (BYOK)

OpenAIAnthropicGeminiOllama

They generate text, answer questions, create drafts

↓

Medha (on-device · Rust · <10ms · $0)

Claim extractionConflict detectionConsensus synthesisEval scoring

↓

You

Decisions informed by measured, traceable evidence

Capabilities

What Medha does - specifically.

Retrieval Evaluation

Integrated

Medha computes deterministic information-retrieval metrics for Korvo’s versioned local evaluation harness without sending private corpus content to a remote evaluator.

Recall@K, Precision@K, MRR, and nDCG

Checksummed datasets and reproducible reports

Exact and native metric implementations

Content-free JSON and Markdown exports

Hard scope-leak constraints

Lexical Baselines

Integrated

Deterministic lexical scoring provides a fast development baseline and candidate-matching primitive. It is a measurement tool, not a claim of semantic correctness.

Local TF-IDF and token-based comparison

Deterministic development baselines

No model calls or per-token cost

Never used as the sole semantic-quality gate

Quality-Gated Experimentation

Harness

Korvo can compare captured retrieval runs against a baseline. Promotion remains blocked until a private, permissioned, human-labelled held-out corpus is approved.

medha_eval_claims - precision, recall, F1 against gold-standard claims

medha_eval_retrieval - nDCG, MRR, Precision@K, Recall@K

medha_eval_coverage - reference claim coverage via sliding-window TF-IDF

medha_suggest_params - next-experiment suggestions via explore/exploit balancing

The Math

Measure changes before
you promote them.

The local harness compares retrieval runs reproducibly. Automated optimization remains gated until held-out data and safe runtime controls are ready.

Cloud LLM Eval

Medha (on-device)

Cost per eval

~$0.03–0.10

Included with Korvo

Time per eval

3–10 seconds

<10 milliseconds

100 iterations

$3–10 · 5–17 min

$0 · 1 second

Internet required

Yes

Works offline

Yes

Data leaves device

Yes

Never

Performance figures depend on operation, device, and dataset size. Deterministic speed does not replace semantic evaluation or human review.

Technical Specs

Medha v1 - what ships today.

Native Rust library, compiled per platform, integrated via Dart FFI.

Language

Rust (native, compiled per platform)

Platforms

macOS, Windows, Linux, iOS, Android

Integration

Dart FFI - C function interface from Flutter

Binary size

~2 MB

Average latency

<10ms per operation

Dependencies

Self-contained - no runtime, no Python, no JVM

Network calls

Zero. Medha never touches the internet.

GPU

Not required - runs on CPU

Cost

Included with Korvo. No per-token billing.

Future Exploration

Roadmap

Medha beyond Korvo.

Medha ships bundled inside Korvo today. External APIs or model releases are not committed until the in-product evaluation workflow is benchmarked and mature.

Medha API

Call Medha's consensus synthesis, claim evaluation, and retrieval scoring from your own applications. REST API with sub-100ms response times.

Consensus synthesis endpoint

Claim extraction & comparison

Eval scoring (nDCG, MRR, F1)

Batch evaluation for pipelines

POST /v1/consensus/synthesize
POST /v1/eval/claims
POST /v1/eval/retrieval

HuggingFace

We're training and refining Medha's evaluation models and plan to publish weights on HuggingFace - so researchers can run, fine-tune, and benchmark independently.

Pre-trained evaluation weights

Claim extraction model

Consensus scoring model

Research-friendly licensing

🤗 korvo/medha-eval-v1
🤗 korvo/medha-consensus-v1
🤗 korvo/medha-claims-v1

Interested in early API access or research collaboration? Get in touch

Honest Positioning

What Medha is - and isn't - today.

✓A deterministic reasoning engine for verification and evaluation

✕A chatbot or text generator (today)

✓Complementary to OpenAI, Anthropic, Gemini - it verifies what they produce

✕A replacement for your AI models

✓Required: zero. Korvo works perfectly without it via cloud synthesis

✕A dependency - it's an accelerator

✓Evolving - deeper reasoning capabilities on the roadmap

✕Staying still - Medha gets smarter with every Korvo release

FAQ

Questions we get asked.

Is Medha an LLM?

Medha is a native deterministic engine for retrieval metrics and fast lexical evaluation. It does not generate text, replace human review, or prove that a claim is true.

Why not use GPT or Claude as the evaluator?

Deterministic metrics are reproducible, inexpensive, and private, so they are useful for retrieval measurement. They are not sufficient for semantic quality; calibrated model-based and human review remain explicit release gates.

What if Medha isn't available on my platform?

Korvo reports native capability availability explicitly. Missing native metrics must not be hidden behind a supposedly equivalent fallback.

Does Medha send any data anywhere?

No. Medha is a compiled native library that runs entirely on your CPU. It makes zero network calls. Your data never leaves your device. This is an architectural guarantee, not a policy checkbox.

Will Medha be available outside of Korvo?

No external release is committed. A standalone interface may be explored after the in-product evaluation workflow and privacy requirements are proven.

Is Medha open source?

Not yet. The core engine (libmedha_ffi) is proprietary and ships bundled with Korvo. We're evaluating open-sourcing components - particularly the eval functions - as Medha matures.

Will Medha ever generate text?

Medha's capabilities will expand. Today it verifies and evaluates. The roadmap includes deeper reasoning capabilities. We'll share more when we're ready.

Already in the app.
Nothing to switch on.

Deterministic on-device evaluation that never touches the internet. Bundled with Korvo and developed behind explicit quality gates.