IndiaAI Mission · Deep Tech Computing
Foundational LLM for Bharat
भारत का
अपना AI
A sovereign, multilingual, multimodal Large Language Model — built in India, for India. Trained on 10+ Indian languages with Sanskrit at its grammatical core, powered by a sparse Mixture-of-Experts architecture.
Sparse MoE Architecture
◆
Sanskrit Grammar Core
◆
Neural Long-Term Memory
◆
CUDA-Optimized Inference
◆
50B Parameter Model
◆
16 Expert Modules
◆
Data Sovereignty
◆
IndiaAI Mission
◆
Sparse MoE Architecture
◆
Sanskrit Grammar Core
◆
Neural Long-Term Memory
◆
CUDA-Optimized Inference
◆
50B Parameter Model
◆
16 Expert Modules
◆
Data Sovereignty
◆
IndiaAI Mission
◆
Hindi
Telugu
Tamil
Gujarati
Bengali
Marathi
Kannada
Odia
Why India Needs This
Data Sovereignty is Non-Negotiable
As a $5 Trillion economy, India cannot afford to have its intellectual capital, governance data, and citizen interactions processed by foreign AI systems. Every query to a US-based LLM is a data sovereignty risk.
Privacy & IP Protection — All data stays within India, processed under Indian law.
22+ Scheduled Languages — Foreign models don’t serve Odia, Santhali, or Konkani speakers well.
Cultural Intelligence — Sanskrit’s grammatical influence gives our model unique cross-language capability.
High-End Jobs in India — Building world-class ML research positions in Bhubaneswar, not Silicon Valley.
Technical Architecture
How the Model Thinks
From raw multilingual input to refined, grammar-validated output — every token routes through specialists, not generalists.
Input
Text, Audio, or Video in any Indian language
→
Preprocessing
Language detect → Tokenization → Normalisation → Morphological analysis
→
MoE Router
Probabilistic gating → Top-K expert selection (K=1–3 of 16)
→
Expert Modules
Sanskrit, Hindi, Math, Code, General, Tamil, Bengali, Odia…
→
Output
Grammar-validated, syntax-corrected response in input language
Core Innovations
Technology That Moves The Needle
Six deep architectural decisions that differentiate this model from generic transformer deployments — each engineered specifically for India’s linguistic and computational reality.
01
Sanskrit Grammar Core
All Indian languages descend from Sanskrit. Our Sanskrit Expert encodes grammar and syntax that generalises across Hindi, Odia, Bengali, Tamil and more — drastically reducing per-language training data needs.
02
Sparse Mixture of Experts
50B total parameters — but only 5–10B activate per query. The MoE router selects the top 1–3 experts from 16 specialised sub-networks, cutting FLOPS by 80% versus dense models like GPT-3.
03
Neural Long-Term Memory
Inspired by the Titans architecture — persistent memory buffers, dynamic updates at inference, and hierarchical attention supporting 2M+ token contexts. No more “getting lost” in long conversations.
04
CUDA & PTX Optimisation
Custom PTX kernels for Devanagari morphological analysis, SIMD/OpenMP parallelism for matrix multiplications, and kernel fusion beyond CUDA Graphs — targeting a 10–30% speedup on H100 GPUs.
05
Task-Based Routing
Rather than language-first routing, we route by task: syntax, semantics, math, code, named-entity recognition. This handles code-switching (Hinglish), low-resource languages, and multi-domain queries gracefully.
06
Diffusion for Generation
A 5B-parameter diffusion expert handles image and video generation via iterative denoising — useful for design, agriculture visualisation, and educational content creation in Indian languages.
MoE Deep Dive
The Router that Never Sleeps
Input Tokens
ସମାଧାନ
x²
–
4
=
0
↓ Router scores all 16 experts ↓
Math
p=0.9
p=0.9
Sanskrit
p=0.7
p=0.7
Code
p=0.2
p=0.2
General
p=0.1
p=0.1
Odia
p=0.4
p=0.4
Math + Sanskrit activated → സമാധാനം: ±2
Why Sparse MoE?
80% Fewer FLOPS.
Same Intelligence.
Dense models like GPT-3 fire every parameter for every input — massively wasteful. Our sparse setup activates only 5–10B of 50B parameters per query, routing to just 1–3 of 16 experts via a learned gating network.
Compute Reduction
80%
KV Cache Reduction (MLA)
93%
Training Iteration Reduction (MTP)
75%
Inference Speed Gain (GQA)
3×
Multilingual Coverage
Languages India Speaks
Initially supporting 4 languages (English, Hindi, Sanskrit, Odia), rapidly expanding to all 22 scheduled languages of India. Each with dedicated token budgets and expert modules.
High
हिं
Hindi
100B tokens target
High
বাং
Bengali
80B tokens target
High
தமி
Tamil
70B tokens target
High
తెలు
Telugu
70B tokens target
Core
संस्
Sanskrit
Grammar engine
High
ଓଡ଼ି
Odia
50B tokens target
Med
ਪੰਜਾ
Punjabi
50B tokens target
Med
ಕನ್ನ
Kannada
60B tokens target
Med
മലയ
Malayalam
60B tokens target
Soon
+13
More Languages
Scheduled expansion
Applications
Built for Real India
Agriculture
Personalised crop management, irrigation, pest control advice in local languages. Voice-based support for smallholder farmers from sensor and satellite data.
Healthcare
Clinical decision support combining modern medicine with digitised Ayurveda (Charak Samhita, Sushruta Samhita). Multilingual patient interaction.
Judiciary & Law
Automated legal research, case summarisation, brief generation. Predict outcomes from historical data. Reduce paperwork and delays in Indian courts.
Education
Multilingual tutoring for JEE, NEET and regional curricula. Offline edge deployment on Jetson Orin Nano for rural schools without internet.
Defence & Security
Intelligence analysis, battle simulation, secure communication, logistics — trained on sensitive military data in an air-gapped OpenStack data centre.
Finance & Accounting
Automated financial reporting, anomaly detection, compliance under RBI guidelines. Document processing in Indian languages.
Research (Co-Scientist)
Literature review automation, hypothesis generation, interdisciplinary connections. Deep research across biology, chemistry, and engineering domains.
Agentic Workflows
MCP-based agents that interact with MATLAB, Adobe Suite, VS Code, and ERP systems — completing complex multi-step tasks autonomously.
12-Month Roadmap
From Zero to 1 Million Users
Phase 1 · Months 1–2
Data Collection & Preprocessing
Collect 100B tokens across Hindi, Sanskrit, Odia, English. Government partnerships with MEITY, state IT departments. PII removal, BPE tokenisation, data augmentation. Launch crowdsourcing pipeline.
₹1.5 Cr
Phase 2 · Months 3–8
Model Training on IndiaAI GPUs
Train 13B–50B parameter sparse MoE model on 8× A100 GPUs via IndiaAI cloud. Target 85% accuracy on core languages. Diffusion expert training for image/video output.
₹3–4 Cr
Phase 3 · Months 9–12
RLHF Fine-Tuning & Pilot Deployment
Fine-tune with 100K+ user interactions. Deploy pilots for 100,000 users in agriculture and healthcare. Optimise inference to ₹0.01/query on 2× L40 server + 1,000 Jetson Orin Nanos for rural edge.
₹3.25 Cr
Infrastructure
Lean, Mean, Indian-First
Training Cluster
NVIDIA A100 / H100 · IndiaAI Cloud
GPUs
8× A100 80GB
GPU Rate
₹150/hr
Training Cost
₹4L / run
Bandwidth
141 GB/s HBM3
Interconnect
NVLink + InfiniBand
Server Inference
NVIDIA L40 · On-Prem / Cloud
GPUs
2× L40 48GB
Throughput
150 q/sec
Latency
<200ms
Cost/Query
₹0.01
Infra Cost
₹15L on-prem
Edge Deployment
Jetson Orin Nano · Rural India
Device
Jetson Orin Nano
Memory
8GB RAM
Model Size
4-bit, ~3GB
Power
15W (solar OK)
Pilot Cost
₹3 Cr / 1000 units
The Team
Small Team, Enormous Ambition
👨💻
Rudhisundar Beura
Co-Founder & Lead Architect
18 years in software, hardware & open source. ML proposals for iDEX military LLM. Built Surabhi.io Social Commerce with AI.
LinkedIn
🌸
Ranjika Beura
Co-Founder & Language & Partnerships Lead
MA in Odia (Sanskrit optional), MBA Finance. Odia & Sanskrit language expert driving data collection and regional language strategy. Leads Government of Odisha relations and private sector client partnerships.
LinkedIn
🧪
Aaditya Kumar
ML Research Engineer
Computational linguistics, LLM efficiency, diffusion models. Key contributor to iDEX military LLM engineering.
LinkedIn
☁️
Aryan Negi
Backend & Cloud Engineer
Spring Boot microservices, Kafka, Kubernetes on Azure. CI/CD automation and scalable API design.
LinkedIn
🏗️
Suvendu Barik
Cloud Architect
Azure/AWS/GCP enterprise architect. Digital banking, cloud migration, IAM, AI/ML platform design.
LinkedIn
Financial Overview
₹15–18 Crore to Build India’s LLM
#
Phase
Scope
Cost (INR)
P1
Data Collection & Preprocessing
Acquisition, cleaning, PII removal, augmentation
₹1.5 Cr
P2
Software Design & Development
Architecture, MoE router, expert modules, APIs
₹4.5 Cr
P3
Model Training & Fine-Tuning
A100 cloud compute, RLHF, diffusion expert
₹3–4 Cr
P4
Inference Optimisation & Deployment
L40 servers, Jetson edge, quantisation
₹1.5–2 Cr
P5
Maintenance, R&D & Expansion
New languages, domain experts, research
₹3 Cr
P6
DevOps, DevSecOps & MLOps
CI/CD pipelines, monitoring, security
₹3 Cr
Total Estimated Budget
₹15–18 Cr
Partner With Us
Help Build India’s Sovereign AI
Whether you’re an investor, government body, academic institution, or enterprise — join us in building the foundational AI that will serve a billion Indians.
A product of Deep Tech Computing Pvt. Ltd. · Bhubaneswar, Odisha, India · Est. 2021