NammaGuide
Unified guidance for Indian public-service processes
NammaGuide AI Lab
Public-interest AI that can be inspected, measured, and improved.
NammaGuide is a Bengaluru-built civic guidance platform exploring how open models, agentic RAG, and local accelerated computing can make complex public processes easier to understand—without pretending AI is an authority.
Founder-built platform
One small local lab. Four assistant experiences. A larger public mission.
Founder Manoj Mukherjee currently develops and operates the private AI gateway on a 32 GB M1 Pro MacBook. That constraint has encouraged efficient routing and evaluation; DGX Spark access would let the lab test more capable local models, parallel agents, and reproducible workshops on a purpose-built NVIDIA platform.
SIR Assistant
A topic-isolated guide for voter-roll revision questions
Astra
A founder-built assistant in the wider AI product portfolio
IVA
A founder-built conversational AI experience
Current baseline
32 GB Apple M1 Pro, an OpenAI-compatible multi-model gateway, grounded retrieval, privacy controls, caching, and observability.
AI Architect Talks
A founder channel for practical architecture, open learning, and conversations that connect builders, students, startups, and enterprises.
Visit the podcast channelResearch direction
A planned doctoral path in applied Quantum AI, alongside near-term research in trustworthy government and civic AI for common Indian service challenges.
Evidence you can inspect
The proposal is backed by live systems, field notes, and source code.
These links let a technical reviewer move from the public API to the architecture decisions and then into the repositories. No private credentials, user data, or protected infrastructure are exposed.
Designed for a fast technical review before a partnership conversation.
Live infrastructure
OpenAI-compatible AI Gateway
Inspect the running API contract and the open-source FastAPI gateway that routes streaming requests to local Ollama models while PostgreSQL, Redis, and Qdrant remain private.
Architecture evidence
Research notes and delivery cases
Review the home-lab topology, agent runtime, Graph RAG experiments, operating tradeoffs, and case studies drawn from real architecture work.
Open-source portfolio
Reusable platform experiments
Explore the founder portfolio and a future-facing micro-frontend proof of concept for independently deployable, AI-native application surfaces.
Multi-agent civic AI
A traceable path from a citizen question to a safer answer.
The goal is not an autonomous authority. It is a coordinated system that selects the right topic, retrieves the right evidence, applies privacy and safety checks, and shows people where to verify the result.
- 01
Route
Detect intent and isolate the service domain
- 02
Retrieve
Search NammaGuide, its graph, and public sources
- 03
Reason
Use the most suitable model lane and specialist agent
- 04
Guard
Check topic fit, privacy, citations, and uncertainty
- 05
Learn
Trace quality signals without storing private questions
Why DGX Spark changes the scale
From constrained experiments to a community AI lab.
NVIDIA lists DGX Spark with a Grace Blackwell GB10 Superchip, 128 GB coherent unified memory, and up to 1 PFLOP FP4 theoretical performance with sparsity. NVIDIA states it can test inference with models up to 200 billion parameters. Actual results depend on model, precision, context, software, and workload.
Read official DGX Spark data32 GB
current unified-memory baseline
128 GB
DGX Spark unified memory
1 PFLOP
up to FP4 with sparsity
200B
up to model size for inference testing
Civic AI research
Evaluate multilingual, source-grounded assistants for common Indian public-service journeys.
Applied Quantum AI horizon
Build the classical AI foundation now for a responsible, evidence-led future doctoral research program.
Open community multiplier
Turn experiments into reusable code, benchmarks, architecture talks, workshops, and learning material.
What exists today
A production-shaped system, not a hardware wishlist.
The collaboration case begins with software already running in the product and a roadmap that can produce reusable evidence for developers, startups, and public-interest technologists.
Grounded retrieval
Every accepted question combines NammaGuide content, a topic-isolated knowledge graph, and allowlisted official-source discovery before synthesis.
Resilient model routing
A private three-model gateway remains primary. NVIDIA Nemotron 3 Ultra is configured through OpenRouter only as the final retryable-failure fallback.
Privacy by design
Sensitive-number detection blocks external calls. Tracing is fail-closed and records operational metadata—not user questions or generated answers.
Proposed DGX Spark validation
A focused 90-day collaboration.
DGX Spark would be used as a local research and demonstration platform. The plan deliberately targets a Spark-suitable Nemotron model for local inference; the much larger Nemotron 3 Ultra integration remains a hosted fallback and is not presented as a local Spark workload.
Public findings would separate measured results from projections and disclose hardware, model, quantization, context, and test conditions.
- 01
Days 1–30
Reproducible local baseline
Deploy a DGX Spark-suitable Nemotron model locally, document the serving stack, and benchmark first-token latency, throughput, memory use, groundedness, and failure recovery.
- 02
Days 31–60
Agentic RAG evaluation
Compare local and hosted inference across NammaGuide’s public-service domains using topic-isolation, citation validity, privacy, latency, and human-review scorecards.
- 03
Days 61–90
Open reference release
Publish a reference architecture, benchmark report, implementation notes, and a Bengaluru developer session showing a production-shaped public-interest AI workflow.
Concrete outputs
What NVIDIA and the community would receive.
- Open-source NVIDIA-ready AI gateway adapter and deployment notes
- Nemotron agentic RAG reference implementation with safety guards
- Reproducible benchmark methodology and public results
- Architecture guide covering local-to-cloud model routing
- Technical article, demo video, and Bengaluru community workshop
- Quarterly impact note covering usage, quality, and lessons learned
From lab to community
Help shape the first NammaGuide AI Community meetup in Bengaluru.
A proposed 10–20 person working session to demonstrate this lab, inspect the architecture, frame useful India-focused AI problems, and start small open artifacts. Participant and sponsor interest is now open.
Direct collaboration inquiry
Start with a clear email—not a blank inbox.
Share what you would like to discuss. AI can tighten the subject and message, while you remain in control of the final wording.
Personal fields are not sent to the AI refiner. Your final inquiry is delivered privately to Manoj’s inbox, and replies go directly to your email.
Primary sources
Official NVIDIA resources
Product and program facts should be checked at NVIDIA. These links are informational and do not indicate a commercial relationship.
NammaGuide is an independent initiative. No NVIDIA sponsorship, partnership, or endorsement is claimed.
NVIDIA, the NVIDIA logo, DGX, Nemotron, and other NVIDIA marks are trademarks or registered trademarks of NVIDIA Corporation in the United States and other countries. The NVIDIA logo is intentionally not displayed pending written authorization.
“Open AI community” describes an open developer community and does not imply affiliation with OpenAI.