Polyglot Swarm

Polyglot Swarm is “a tutor that remembers, not a chatbot that forgets”. A LangGraph graph routes each turn through a conversation agent and a parallel fan-out of grammar, vocabulary and cultural agents, then an evaluator reconciles their findings before the loop continues. Underneath, an evidence pipeline turns every exchange into learning events that update a transparent per-skill learner model.

Language-practice chatbots forget everything between sessions and confidently correct things that were right. A useful tutor has to remember, model what the learner can actually do skill by skill, and abstain when it is not sure — a false correction is treated as worse than a missed one.

The project's own implementation plan is explicit that the core is the learner model, not the number of agents: build the evidence, verification and evaluation layers rather than adding more personas.

Router → conversation → parallel grammar, vocabulary and cultural agents → evaluator → router loop, with CEFR assessment and cross-language transfer computed at session end.

Evidence pipeline

Extractor → normaliser → deduplicator → store of LearningEvents, persisted per user, language and session, then fed to a belief layer.

Learner model

A MasteryEngine that is deliberately a transparent weighted-evidence rule — not a Bayesian network, IRT or a trained model — over seven skills (speaking, listening, reading, writing, grammar, vocabulary, pragmatics) with mastery, uncertainty and sample size.

Verifier

The evaluator accepts, revises or abstains on proposed corrections, reconciles grammar-versus-culture conflicts, and its calibration is measured with expected calibration error.

Review scheduling

FSRS via py-fsrs decides what to review and when.

Voice

Whisper for speech-to-text and edge-tts for speech, with a VAD module.

Reliability

A routing provider with tiered fallback, exponential backoff and a circuit breaker; if every provider fails the tutor still replies in character.

Storage and surfaces

SQLite as the source of truth with additive migrations, ChromaDB collections for vocabulary, grammar rules, conversations and cultural notes; FastAPI, Gradio, a Next.js dashboard and a Typer CLI.

• Python 3.12 with LangGraph, LangChain provider packages, ChromaDB, FastAPI, Gradio, pydantic-settings and a versioned Jinja prompt registry; ruff, strict mypy, pytest and Hypothesis.
• Thirty architecture decision records document the trade-offs.
• A deterministic fake-LLM mode makes the whole suite and the benchmarks runnable offline.

1. 01

   Transparent over trained

   Mastery is computed by an inspectable weighted-evidence rule so every score can be explained back to the events that produced it.

2. 02

   Abstain when unsure

   False corrections are treated as worse than missed ones, so uncertain calls abstain rather than guess.

3. 03

   Learner model first

   Investment goes into evidence, verification and evaluation rather than into more agents.

Benchmarks

Grammar precision · verifier calibration (ECE) · structured output · provider latency

Ablation

5 arms

Conversation only through the full stack, run on scripted learners in deterministic mode.

Regression evals

Grammar, vocabulary, assessment and curriculum suites

*All reported numbers come from deterministic fake-LLM mode; there has been no study with real learners.

• Keeping fourteen-plus phases of features additive without breaking the graph's state shape or test-pinned behaviour.
• Calibrating a verifier that must prefer silence to a confident wrong correction.

• No study with real learners; CEFR estimates are not validated against expert human raters.
• Text-to-speech uses a network service, so “fully local” applies to text and Whisper transcription.
• Alpha software.

• Evidence first, beliefs second, learner model third — the ordering of persistence determines whether anything downstream is trustworthy.
• An abstaining verifier is a product decision as much as a modelling one.

• Repository (opens in a new tab)github.com/NiravRVaghasiya/Polyglot-Swarm
• Architecture (opens in a new tab)github.com/NiravRVaghasiya/Polyglot-Swarm/blob/main/docs/architecture.md
• Learner model (opens in a new tab)github.com/NiravRVaghasiya/Polyglot-Swarm/blob/main/docs/learner-model.md
• Decision records (opens in a new tab)github.com/NiravRVaghasiya/Polyglot-Swarm/tree/main/docs/adr