GIN AI Research Lab | gin.info.vn — Applied AI, Sovereign Edge & Human-Centric Systems

Core Research Vectors

Four Pillars of Sovereign Intelligence

We solve fundamental engineering challenges: drastic compute footprint reduction, absolute data sovereignty, and aligning autonomous agents with practical human workflows.

1. Extreme Low-Bit & Edge AI

Training from-scratch 1.58-bit ternary architectures with Quantization-Aware Training (QAT) and Straight-Through Estimators (STE). Achieving near-lossless machine translation at ~350 tok/s on standard laptop CPUs with a 77.56 MB disk footprint.

Project: Bit-Translate

2. Full-Duplex Multimodal Agents

Engineering synchronous browser-based classrooms with sub-second voice activity interruption (Barge-in VAD), interleaved pedagogical reasoning, and autonomous DOM tool calling that dynamically drives visual slides, lip syncing, and stroke animations.

Project: AI Live Sensei

3. AI FinOps & Agent Governance

Deep lifecycle hooks telemetry and cost analytics built for autonomous coding agent fleets. Quantifying Prompt Caching savings in USD, visualizing tool execution flows via Sankey graphs, and enforcing zero-prompt privacy.

Project: KS-Dashboard

4. Sovereign Digital Health

Evidence-informed maternal and child health research. 100% offline-first PWA, clinical growth percentile calculators, CDC milk-stash management, Edinburgh Postnatal Depression Scale (EPDS) screening, and zero third-party tracking.

Project: GinBaby

Live Projects & Prototypes

Active Research Portfolio

Every system is an open-source, verifiable, production-grade working implementation.

5 Active Systems All Code Open & Tested

EdTech Multimodal AI 110 Lessons Kana→N1 Full-Duplex VAD

AI Live Sensei Classroom

An autonomous browser-based classroom delivering synchronous 1-on-1 human-grade language tutoring. Features bidirectional streaming audio (16kHz in, 24kHz out), sub-second natural interruption (Barge-in VAD), interleaved pedagogical reasoning stream, and autonomous DOM tool-calling agent (change_slide, highlight_element, mark_error). Over 1,500+ generated lesson illustrations and 28 lip-synced characters.

Turnaround VAD

< 1.0s Interruption

Curriculum Scale

110 Lessons (N5–N1)

Visual Pipeline

Automated 1080p Video

Explore Repository MIT / Custom License

Live Stage Engine (Bảng đen)

Extreme Edge AI 1.58-Bit Ternary Weights No GPU Required

Bit-Translate: 1.58-Bit Neural MT

From-scratch 1.58-bit ternary weight ({-1, 0, +1}) Neural Machine Translation (Japanese ↔ Vietnamese) based on the BitNet b1.58 architecture. 152.1M parameters trained on 15.98M parallel pairs and compressed down to a 77.56 MB GGUF binary. Runs at ~350 tokens/sec on standard laptop CPUs with only ~125 MB working RAM.

Model / System	Accuracy (Score 2)	Speed (CPU)	Size / RAM
Bit-Translate v7a i2s	78.5%	~350 tok/s	77.56 MB / ~125 MB
Google Translate Cloud	69.0%	Cloud API Latency	Cloud Bound

Explore Repository GGUF i2_s Deployable

Verified Physical Profiling

• Layers: 18 (d_model=768, FFN=2048, 12 heads)

• Training: 15,986,646 sentence pairs (951M tokens)

• Quantization: QAT + Straight-Through Estimators

• Blind Judge: Opus 5 evaluation on 200 held-out pairs

• Difficult Idiom Set: 39W - 7L (85% win rate)

• Extreme Floor Check: PTQ <2.5 bpw collapses PPL to 11,340

AI FinOps & Observability Agent Telemetry Zero-Prompt Privacy

KS-Dashboard: Developer AI FinOps & Governance

A self-hosted enterprise control center built for autonomous coding agent fleets (such as Claude Code and custom agentic CLI harnesses). Hooks directly into lifecycle events (SessionStart, PreToolUse, PostToolUse, Stop). Tracks token burn, computes real USD savings from Prompt Caching, visualizes tool transition Sankeys, and tracks team adoption—all without ever storing prompt text or corporate source code.

Prompt Cache ROI

Exact USD Computed

Real-Time Monitor

SSE Live Event Bus

Enterprise SSO

Microsoft Entra ID + AMIS

Explore Repository Next.js 16 + Prisma 6

Lifecycle Telemetry Architecture OTEL_LOG_USER_PROMPTS=0


                                        SessionStart ➔ ToolExecution ➔ CostEngine

                                        • Token Input / Output Breakdown

                                        • Cache Creation vs. Cache Read Hits

                                        • Tool Latency Percentiles (P50/P90/P99)

                                        • Task Classification (Debug/Code/Plan)

Guarantees zero corporate intellectual property retention while giving engineering directors complete visibility into model spend.

Sovereign Edge Computing 100% Offline / Air-Gapped Windows Transparent Overlay

TransX: Sovereign Edge Speech Overlay

A desktop translation overlay engineered for high-security, defense, and confidential corporate boardrooms. Completely air-gapped: zero external HTTP calls during inference. Combines local Faster-Whisper (CTranslate2 INT8, 2–4x faster than whisper.cpp) with NLLB-200 INT8 (~594 MB) via an English pivot and contextual sliding-window accumulation.

Audio Capture

WASAPI Loopback

RAM Footprint

~900 MB (Whisper+NLLB)

Security Tier

Zero Telemetry

Explore Repository MIT License

Linguistic Gating Parameters

• Japanese (SOV): chunkMaxMs=10s, silenceMs=600ms

• English/VI (SVO): chunkMaxMs=8s, silenceMs=1000ms

• Syntactic Gating: SentenceAccumulator terminal flush

• Pivot Route: JP ➔ [beam=6] ➔ EN ➔ [beam=4, ctx] ➔ VI

• Hotkey: Ctrl+Shift+T with click-through overlay

Digital Health Research Flutter 3.47 PWA & iOS Offline-First Private

GinBaby: Maternal & Pediatric Companion

A calm, private, evidence-informed parenting and digital health research companion. 15.6k lines of Dart with 97 passing automated tests. Built specifically for pumping mothers (dual-side timer, CDC milk stash, estimated direct latch). Features drag-to-fill bottle with AAP citations, EASY routine nap engine, WHO 2006 growth percentiles, and Edinburgh Postnatal Depression Scale (EPDS) self-screening. Zero ads, zero tracking, all data stays on device.

Test Suite

97 Tests Passing

Clinical Standards

WHO 2006 + CDC + EPDS

Design System

Watercolour Art (~50 assets)

Explore Repository PWA Installable

Flutter PWA Interactive Flow

Engineering Rigor

Our Research Methodology

We hold our work to strict empirical standards: zero metric extrapolation, reproducible test benches, and physical hardware verification.

Zero Extrapolation Policy

Every figure in our papers and repos originates from executed tool runs, physical OS counters, or blinded multi-model audits. Unmeasured benchmarks are strictly labeled as untested.

Sovereign Data Privacy

We design systems where user data never leaves the edge device by default. Our enterprise FinOps and digital health suites enforce strict client-side indexing and zero-telemetry architectures.

Open & Reproducible

All codebases, conversion scripts, synthetic teacher prompts, and testing suites are published publicly on GitHub with complete instructions to reproduce every finding locally.

Future Outlook

12-Month Research Roadmap

Target milestones across model scaling, clinical trials, and multi-agent developer infrastructure.

PHASE 1 (Q1–Q2)

Hosted Tutoring & 300M Model Tier

Scale Bit-Translate architecture to a 300M-parameter tier under a 150MB binary budget. Deploy multi-tenant hosted Sensei Classroom for pilot cohorts preparing for JLPT examinations.

PHASE 2 (Q3)

Multi-Agent FinOps & Automated Breakers

Expand KS-Dashboard adapters beyond CLI to multi-agent harnesses (Cursor, Aider, OpenHands). Release automated budget threshold webhooks and circuit breakers.

PHASE 3 (Q4)

Clinician Review & Sovereign Health Study

Conduct an ethical, consented pilot study with new mothers in Vietnam using GinBaby to evaluate the impact of private postpartum depression tracking and feeding routines.

G

Collaborate with GIN AI Research

Whether you are a researcher, open-source contributor, academic institution, or technology partner, we welcome inquiries, peer discussions, and collaborative engineering.

github.com/GIN-AI-Research-Lab • gin.info.vn • tuent@gin.info.vn

Pioneering Extreme Edge AI, Full-Duplex Agents & FinOps.