Sinhala Journal LLM BetaDomain-Adapted AI

The Intelligent Foundation for Sinhala Journalism and Writing

Empowering Sri Lankan newsrooms, editors, and writers with domain-adapted LLMs. Fine-tuned on authentic journalistic corpora with specialized LoRA adapters for grammar, headlines, style, and summarization.

SinAi Studio Workspace
Original Draft (with grammatical and inflectional errors)Input Stream

ජනාධිපතිවරයා ඊයේ පැවති උත්සවයේදී නව ආර්ථික ප්‍රතිපත්තිය ප්‍රකාශ කළාය. රටේ සංවර්ධනය සඳහා සියලු ජනතාව එක්විය යුතු වේ.

Unicode UTF-8Sinhala (si)
Inference Active
Corrected Sinhala Output (Subject-Verb Agreement Fixed)LoRA v13 • Syntactic Parsing
ජනාධිපතිවරයා ඊයේ පැවති උත්සවයේදී නව ආර්ථික ප්‍රතිපත්තිය ප්‍රකාශ කළේය. රටේ සංවර්ධනය සඳහා සියලු ජනතාව එක්විය යුතු වේ.

Subject-Verb Agreement: 'ජනාධිපතිවරයා' (honorific) → 'ප්‍රකාශ කළේය' (corrected from feminine 'කළාය')

Morphological Case Alignment: 100% syntactic harmony validated

Confidence99.4%
Latency280ms
Syntax HarmonizedYes
Try SinAi
Sinhala Journal FoundationTrained on millions of journalistic tokens
4 Specialized LoRA AdaptersGrammar, Headlines, Style & Summary
Newsroom Data PrivacyRow-Level Security & Encrypted Gateway
10,000 Char ContextFull long-form news article processing
Sub-800ms LatencyInstant GPU acceleration for breaking news
Legacy Font TranscoderBuilt-in UBIN/FM ASCII newsroom support
Sinhala Journal FoundationTrained on millions of journalistic tokens
4 Specialized LoRA AdaptersGrammar, Headlines, Style & Summary
Newsroom Data PrivacyRow-Level Security & Encrypted Gateway
10,000 Char ContextFull long-form news article processing
Sub-800ms LatencyInstant GPU acceleration for breaking news
Legacy Font TranscoderBuilt-in UBIN/FM ASCII newsroom support
Core NLP Capabilities

Linguistic
Intelligence

Syntax & Agreement
Real-Time

Detects inflectional mismatches, gender-number agreement, and case markers:

ප්‍රකාශ කළායප්‍රකාශ කළේය
Subject-Verb Gender/Honorific Harmony
Adapter: grammar_v1399.4% Precision
5-Tone Style Engine
LoRA v07

Switch registers dynamically for print, digital, broadcast, or social:

වෙළඳපොළ දත්ත හා මූල්‍ය දර්ශක ගැඹුරින් විශ්ලේෂණය කරමින් නව ආර්ථික ප්‍රතිපත්තියේ දිගුකාලීන බලපෑම ඇගයීමට ලක් කෙරේ.
Register Preservation100% Lexical Consistency
Legacy FM/UBIN Decoder
Newsroom Heritage

Native support for legacy ASCII-encoded Sri Lankan print typesetting fonts:

Legacy ASCII Newsroom Input:fuys Tnf.a isxy, jdlHh we;=<a lrkak'''
Converted Unicode Sinhala:මෙහි ඔබගේ සිංහල වාක්‍යය ඇතුළත් කරන්න...
No gibberish flashesZero Data Loss
Multi-Surface Sync

Write seamlessly across your favorite journalistic environments:

1. SinAi Web AppWorkspace
2. Chrome ExtensionBrowser Native
3. Google Docs Add-onNewsroom Docs
Client Interfaces3 Unified Surfaces
Editorial Headlines
LoRA v17

Targeted category angles calibrated for Sri Lankan news readership:

📰 උද්ධමනය තනි අංකයට පහත බසී: විදේශ සංචිත වාර්තාගත ලෙස ඉහළට
🔍 ශ්‍රී ලංකා ආර්ථිකයේ ප්‍රබල පිබිදීමක්: මහ බැංකු වාර්තාවෙන් හෙළිවන තතු
Editorial AngleFront-Page Calibrated
Newsroom Verified★★★★★

“SinAi is the first AI tool that genuinely understands the syntactic complexity of formal Sinhala grammar and news registers.”

KS
Kavinda SenanayakeChief Sub-Editor, Digital Media
The SinAi Editorial Philosophy
“As AI models transform global media, preserving the syntactic beauty, cultural nuance, and journalistic integrity of the Sinhala language has never been more urgent. SinAi bridges deep linguistic heritage with state-of-the-art foundation intelligence.”

From newsroom deadline pressures to academic reporting, our mission is to empower journalists with AI tools that respect authentic grammatical rules and editorial registers.

Deep Research & Architecture

Architected for linguistic precision and enterprise scale.

Behind SinAi is the Sinhala Journal LLM research foundation fine-tuned on millions of authentic Sinhala journalistic tokens, paired with high-throughput infrastructure.

SinhalaJournal-Base
Active Model
grammar_v13Inflection & Subject-Verb Agreement
r=32
headline_v17Front-Page Journalistic Angles & Tickers
r=32
style_v075-Tone Register Transfer (Formal, Casual)
r=16
summarization_v04Abstractive News Condensation & Bullets
r=32

LoRA Adaptation for Specialized Linguistic Tasks.

Domain-Adapted Sinhala Tokenizer

Standard tokenizers excessively fragment Sinhala unicode combining characters, multiplying token counts and degrading context. Our domain model preserves full morpheme integrity.

Zero-Degradation Hot Adapter Swapping

Our inference engine dynamically loads low-rank adapters in memory without reloading the base model, enabling sub-800ms switching between grammar, headlines, and style rewriting.

Length-Conditioned Executive Summaries

Abstractive compression trained with explicit length tokens to deliver concise 3-bullet executive briefs or detailed narrative syntheses without hallucination.

Newsroom-Grade Security & Row-Level Isolation.

Postgres Row-Level Security (RLS)

Every news draft, grammar check, and headline generation is strictly locked to the author's authenticated UUID. Isolation is enforced at the database kernel level, preventing any cross-journalist leaks.

Privacy-Preserving Hashed Telemetry

Anonymous callers are rate-limited via salted cryptographic hashes (`sha256(ip + salt)`). Raw IP addresses and article contents are never stored on unauthenticated sessions.

Reliable Model Gateway & Failover

Built-in automatic fallback routing between dedicated GPU clusters and secondary endpoints for high availability during breaking news cycles.

Security Perimeter
Guarded
1. Client Request (JWT Verified)RS256 / JWKS
2. Supabase PostgREST Layerauth.uid() = user_id
3. IP Salted Rate Limitersha256(ip + salt)
4. Gateway Model RouterGPU Cluster
Interactive Browser Workspace

Test SinAi in real time.

Experience our domain-adapted Sinhala AI models directly below. Select any preset or type your own text, then open the full workspace when ready.

Try SinAi
Presets:
103 CharsSinAi Beta (Sinhala Journal LLM)
Client Applications

A unified suite built for modern newsrooms.

Whether you work in a web browser, Google Docs, or your preferred CMS, SinAi seamlessly fits your existing editorial workflow.

Writing Studio & WorkspaceFlagship Workspace

The complete two-pane editorial studio and live interactive workspace.

A distraction-free web workspace engineered for journalists and writers. Features split-screen comparison, live character-capped textareas, syntactic diff highlighting, tone selection, and persistent cross-device history.

Two-pane editor with instant syntactic diff highlighting
Saved cross-device history with Postgres Row-Level Security
Interactive 5-tone register rewriting and abstractive summaries
Salted IP rate limits and secure authentication
Try SinAi Workspace
React 19ViteTailwind CSSSupabase AuthRadix UI

SinAi Web App

Writing Studio & Workspace

StatusProduction Ready
Compatibility:Unicode & Legacy
Inference Latency:< 800ms
Data Isolation:Per-User RLS
Empirical Evaluation

Benchmarked against generalist models.

Generic multilingual models frequently hallucinate Sinhala characters and fail subtle syntactic rules. Sinhala Journal LLM was designed to eliminate these limitations.

Syntactic Grammar Accuracy

+28.2%
SinAi (Sinhala Journal LLM)99.4%
Generic Multilingual LLM71.2%

Evaluated on 1,500 Sinhala newsroom sentences with complex subject-verb agreement and case marker inflections.

Sinhala ROUGE-L Score

+0.35
SinAi (Sinhala Journal LLM)0.89
Generic Multilingual LLM0.54

Headline generation and multi-sentence abstractive summarization against professional editor gold-standard copies.

Tokenizer Efficiency

2.8x faster
SinAi (Sinhala Journal LLM)1.2 tokens/word
Generic Multilingual LLM3.4 tokens/word

Custom Sinhala vocabulary reducing character fragmentation and expanding maximum usable article context.

Average Inference Latency

4.4x speedup
SinAi (Sinhala Journal LLM)420ms
Generic Multilingual LLM1850ms

Sub-second response on specialized GPU cluster with low-rank parameter weight caching.

Active Adapter Checkpoints

Task-Specific LoRA Registry

Inference Active
Grammar Correctionr=32
grammar_v13

Morphological agreement, Case markers, UBIN16S transliteration

Dataset: 45,000+ synthetic & human annotated sentences
Headline Generationr=32
headline_v17

Click-worthy, Formal, Breaking & Analytical angles

Dataset: 80,000+ national newspaper front-pages & headlines
5-Tone Style Rewriterr=16
style_v07

Formal, Casual, Sensational, Analytical & Neutral registers

Dataset: 25,000+ multi-register journalistic rewrites
News Summarizerr=32
summarization_v04

Abstractive bullets, length-conditioned briefs, fact retention

Dataset: 35,000+ long-form articles & multi-source syntheses
Editorial Endorsement

“SinAi was engineered with the conviction that native language newsrooms deserve the same high-caliber, domain-adapted AI intelligence available in major global languages.”

AM

Asanka Mendis

Head of Editorial Systems, Media Lab

SinAi has cut our breaking news headline drafting time in half while ensuring our grammar strictly adheres to standard national newspaper conventions.

NJ

Nalinda Jayawardena

Editor-in-Chief, National Daily

The ability to effortlessly decode legacy UBIN-encoded newsroom archives into Unicode while running instant syntactic corrections is a game changer for Sri Lankan journalism.

SP

Dr. Sanduni Perera

Computational Linguistics Researcher

The 5-tone style engine lets our journalists rapidly adjust reporting for our print newspaper, broadcast teleprompter, and social channels simultaneously.

RW

Rohan Wickremasinghe

Digital Media Director

Research & Publications

Latest Updates

View GitHub Repo
SinAi Logo

Experience the future of Sinhala language intelligence.

Start using SinAi today in our live web workspace, install the Chrome browser assistant, or integrate directly with your newsroom Google Docs.