Words Today
28,400
500k/day Target · 8-Core CPU
Total Requests
1,420
24.5 /s burst capacity
Cache Hit Ratio
84.2%
L1 Memory + L2 Redis + L3 SQLite
Latency (P50 / P95)
12.4 ms
P95: 48.2 ms
🖥️ Server Hardware Telemetry
Dedicated 8-Core Intel/AMD CPU with 32 GB RAM
CPU Core Saturation (8 Cores)
18.4%
RAM Consumption (32 GB NVMe System)
24.2% (7.7 / 32 GB)
🟢 CTranslate2 Engine: INT8 Optimized OpenMP (intra_threads=4, inter_threads=2)
🟢 Distributed Redis L2: Connected (172.18.0.4:6379) · Single-Flight Stampede Barrier Active
🟢 Preserved Checkpoints:
lora-indictrans/checkpoint-5 verified intact🛡️ High-Reliability Architecture Status
Production coexistence with co-located apps
Co-located Services (whatsappapi, appAnalytics, image-api, jalan, email)
100% UPTIME
Romanized Indic (Hinglish) Grammar Concord
ACTIVE & VERIFIED
Model Cache Key Isolation (No Bleed)
VERSION-NAMESPACED
Instant Rollback Readiness
< 1 MS ATOMIC SWAP
🌐 Interactive Translation & Hinglish Playground
Bidirectional 22 Indian languages + English with Romanized Indic normalization & loanword gender concord
Ready
Try Presets:
✨ Automatically executes Romanized Indic normalization & loanword gender concord
📁 Translation Catalogs & Fine-Tuning Datasets
Drag & drop CSV, JSON, or JSONL files with automated column schema detection and Unicode verification
Drag & drop dataset file here, or click to browse
Supports .csv, .json, .jsonl · Automatically detects source and target text columns
Active Datasets Catalog
| Dataset Name | Format | Records | Languages | Columns | Created | Action |
|---|
🎯 Launch Fine-Tuning Job
Trains PEFT LoRA adapter in an isolated background thread without degrading production API
Fine-Tuning History
| Job ID | Dataset | Architecture | Status | Progress | Final Loss | Started |
|---|
🧠 Versioned Model Registry
Tracks model versions, parameters, quantization, status badges, and 13-criteria quality gate metrics
| Model Version | Architecture | Quantization | Status | Traffic Split | Quality (chrF++) | Registered At |
|---|
🚀 Zero-Downtime Canary Traffic Controller
Safely split user traffic between Production and Candidate models without restarting server processes
🛡️ Instant Zero-Downtime Rollback Guard
If any quality anomaly, regression, or error spike occurs on a canary release, triggering rollback instantaneously swaps the routing pointer back to the frozen baseline (Model V1) in under 1 millisecond.
🏅 13-Dimension Production Quality Gate Scorecard
Evaluated against the 1,008-case Golden Dataset across all 22 scheduled Indian languages
📋 Real-Time System Execution Logs
Audit trail, router events, and cache hit telemetry
[INFO] System initialized on 8-Core Ubuntu 24.04 LTS host.
[INFO] CTranslate2 IndicTrans2 INT8 provider loaded in OpenMP multi-thread mode.
[INFO] Redis L2 distributed cache active with Single-Flight Stampede Barrier.
[INFO] Model Registry initialized: Production baseline = model_v1.
[TRANSFORM] req_id=golden-001 | en -> hi | mode=translate | nodes=1 | elapsed=14ms
[HINGLISH] Grammatical concord applied: 'kya tractor ka service ho gaya hai?' ➔ 'क्या ट्रैक्टर की सर्विस हो गई है?'
[INFO] Nightly batch scheduler configured for 23:00 daily.