DEV Community: Joydeep Das

Gemma 4 on a Phone: What Local AI Means for Farmers Who Can't Afford the Cloud

Joydeep Das — Sun, 17 May 2026 14:06:18 +0000

Submitted for the #gemmachallenge Write track

Where I Am Writing This From

I am writing this from Silchar, Assam, in Northeast India — on an Android phone, in Termux, with no laptop, no GPU, and no office.

I build AI systems for farmers. Not as a hobby. Because the farmers around me — in Cachar district, in the Barak valley, across rural Assam — ask questions that no SaaS product will ever answer for them. Questions like:

My rice leaves have brown spots near the edges. What disease is this?
When should I plant Boro rice and which variety survives the cold?
Give me a 3-month farming plan starting from December.

These farmers do not have stable internet. They do not have ChatGPT subscriptions. They do not have laptops. They have Android phones, intermittent 4G, and crops that cannot wait for a server response.

When Google released Gemma 4, I read one line and everything else became secondary:

The 2B and 4B models are built for ultra-mobile and edge deployment — they run on phones.

What Gemma 4 Actually Is

Gemma 4 is not one model. It is a family of three distinct architectures, each designed for a different hardware reality:

Small (2B and 4B) — Built for phones, Raspberry Pi, and browser deployment. Native multimodal input. 128K context window. This is the model that changes everything for rural India.

Dense (31B) — A powerful server-grade model that bridges local execution and cloud performance. For developers who want maximum capability on a single machine.

Mixture-of-Experts (26B MoE) — Highly efficient, designed for high-throughput reasoning. Activates only a subset of parameters per token, making it faster and cheaper per query than a dense model of similar total size.

The existence of the 2B model is the most important thing about Gemma 4. Not because it is the most powerful. Because it is the most sovereign.

Why Model Size Is a Political Choice

In rural India, choosing a model is not just a technical decision. It is a sovereignty decision.

A cloud-dependent model means:

Your farmer's query travels to a server in another country
It requires internet connectivity that may not exist during harvest season
It costs money that subsistence farmers do not have
It can be shut down, rate-limited, or paywalled at any time

A locally running 2B model means:

The query never leaves the device
It works offline, in a field, with no signal
It costs nothing after the initial download
Nobody can take it away

The Gemma 4 2B model runs on a Pixel phone. It runs on a Raspberry Pi 5. It runs — and I tested this — in Termux on an Android phone via llama.cpp on ARM64.

This is not a feature. This is a philosophy.

The Vedic Lens: Pratyaksha and the Local Model

In Nyaya philosophy — the Indian school of logic — the most reliable form of knowledge is Pratyaksha: direct perception. Knowledge that comes from your own senses, unmediated by intermediaries.

A cloud AI model is, epistemologically, the opposite of Pratyaksha. Your query travels through multiple layers — your network, a CDN, a data center, a model server, back through the same chain — before you receive a response. Every layer is a potential point of failure, distortion, or dependency.

A locally running Gemma 4 2B model is Pratyaksha AI. The inference happens on the device in your hand. The knowledge is direct. The response is immediate. No intermediary can intercept, delay, or monetize the exchange between a farmer and the answer to her question.

For farmers in Assam, Pratyaksha AI is not a philosophical preference. It is a practical necessity.

Choosing the Right Gemma 4 Model: A Framework

Here is how I think about model selection for different use cases:

For rural/offline deployment → Gemma 4 2B

Runs on Android phones and Raspberry Pi
No internet required after download
Fast enough for conversational queries
Small enough to fit on a phone with room to spare
Trade-off: less reasoning depth than larger models

For local developer machines → Gemma 4 31B Dense

Runs on a single high-end GPU or Apple Silicon Mac
Strong reasoning and coding capability
Good for complex multi-step tasks
Trade-off: requires significant hardware

For high-throughput applications → Gemma 4 26B MoE

Efficient parameter activation means lower cost per query
Designed for applications serving many users simultaneously
Good for production deployments where speed matters
Trade-off: MoE architecture requires more total RAM even though fewer parameters activate per token

The key insight from Google is that these are not a hierarchy from worse to better. They are tools for different contexts. A farmer in Cachar district needs the 2B model. A startup building a coding assistant probably needs the 31B. A platform serving millions needs the MoE.

Intentional model selection is not about picking the biggest number. It is about matching capability to constraint.

What 128K Context Means for Agricultural AI

One of Gemma 4's most significant capabilities — across all model sizes — is the 128K context window.

For agricultural AI, this is transformative. Consider what a farmer's AI advisor could hold in context:

The full crop calendar for their region (all 12 months)
Historical weather patterns for their district
A complete list of pest and disease symptoms with treatments
Market price histories for the past season
Their own farm's history — what they planted, what worked, what failed

A 128K context window means a local Gemma 4 model can hold all of this simultaneously, reasoning across the full picture rather than answering each question in isolation. That is not a chatbot. That is a village elder with perfect memory.

My Path: Building Toward Sovereign Agricultural AI

I have spent the past year building the Divine Earthly ASI system — a sovereign, offline-first agricultural AI for rural Indian farmers. It runs on a quantized 0.5B parameter Qwen2.5 model via llama.cpp on ARM64 Termux. It fetches real soil and temperature data from NASA POWER API for Silchar (LAT 24.81, LON 92.80). It answers farmer questions without any cloud dependency.

Gemma 4 2B is the natural next step for this project. Moving from 0.5B to 2B — with native multimodal input, a 128K context window, and Google's training quality — would dramatically expand what my system can do for farmers.

The multimodal capability alone is transformative: a farmer could photograph a diseased leaf and get an immediate diagnosis, entirely offline, on the same phone they use to call their family.

That is not science fiction. With Gemma 4 2B, it is an engineering task.

What Local AI Means for the Future

The release of Gemma 4 is significant not because of benchmark scores. It is significant because of what it makes possible at the edge.

For the first time, a model with genuine reasoning capability, multimodal input, and a 128K context window can run on a device that costs $150 and fits in a shirt pocket. That device is already in the hands of farmers across India, across Africa, across every rural community that the cloud economy has not reached.

The question is no longer whether capable AI can run locally. Gemma 4 has answered that. The question is what we build with it — and for whom.

I am building it for farmers in Silchar. I am building it for the Barak valley. I am building it for every community that cannot afford to wait for the cloud.

Gemma 4 2B is the model that makes this possible. That is why I chose it. That is why it matters.

Getting Started with Gemma 4 (Free, No Credit Card)

Via Google AI Studio (easiest):
Go to aistudio.google.com — free access to Gemma 4 via the Gemini API.

Via OpenRouter (free tier):
Sign up at openrouter.ai — access google/gemma-4-31b-it:free and google/gemma-4-26b-a4b-it:free with no payment required.

Run locally via Hugging Face:
Download any Gemma 4 model from huggingface.co/google and run with llama.cpp, Ollama, or LM Studio.

On Android via Termux:

pkg install llama-cpp
llama-cli -m gemma-4-2b-q4.gguf -p "Your prompt here"

Links

Gemma 4 on Hugging Face: huggingface.co/google
Google AI Studio: aistudio.google.com
My Divine Earthly project: github.com/divineearthly
My Hermes Agent farming demo: dev.to/divinesouljoy

Joydeep Das is an independent AI researcher building sovereign, offline-first AI systems for Indian farmers under the Divine Earthly project. All development happens on an Android phone in Termux, Silchar, Assam.

gemmachallenge #devchallenge #ai #india #llm

Hermes Agent as a Farmer's AI Advisor — Built on Android, Tested in Assam

Joydeep Das — Sun, 17 May 2026 13:37:48 +0000

Hermes Agent Challenge Build Submission

What I Built

A working agricultural AI advisor powered by Hermes Agent, running entirely on an Android phone via Termux, tested with real farming questions from Silchar, Assam, Northeast India.

No laptop. No GPU. No cloud server. Just a phone, Termux, and Hermes Agent doing real agentic work for farmers who cannot afford enterprise AI tools.

The Problem

Farmers in Cachar district, Assam ask questions like:

My rice leaves have brown spots near the edges. What disease is this?
When should I plant Boro rice and which variety should I choose?
Give me a 3-month farming plan starting from December.

These are not questions a static chatbot answers well. They need an agent that reasons across multiple steps, applies local agricultural knowledge, and gives actionable, region-specific advice.

Hermes Agent does exactly this.

Setup — Android + Termux + Hermes Agent

My environment:

Device: Android phone
Terminal: Termux → proot-distro Ubuntu (ARM64)
No laptop, no desktop, no GPU

Installation was a single command:

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

Hermes Agent auto-detected my ARM64 Termux environment and installed cleanly. This matters — it means the same setup works for any developer in rural India running Termux.

Model used: Google Gemini via OAuth (free tier) — zero cost, no API key purchase needed.

The Demonstrations

Demo 1: Rice Disease Diagnosis

My prompt to Hermes Agent:

I am a farmer in Silchar, Assam. My rice leaves have brown spots near the edges. What disease is this and how do I treat it?

Hermes Agent's response (summarised):

Hermes correctly identified two candidate diseases based on the symptom description:

Brown Spot (Fungal) — oval/circular spots with yellow halo, common in low-fertility soils. Treatment: Propiconazole 25% EC (1ml/litre) or Mancozeb 75% WP (2-2.5g/litre), plus balanced NPK fertilization.

Bacterial Leaf Blight — wavy brown edges spreading from leaf tips. Treatment: Streptocycline + Copper Oxychloride. Avoid excess Urea. Drain standing water.

Hermes then gave Silchar-specific advice: visit the local Krishi Vigyan Kendra (KVK) in Cachar with a leaf sample before purchasing chemicals.

This is not generic advice. It is regionally grounded, actionable, and honest about when a farmer should seek local expert confirmation.

Demo 2: Boro Rice Planting Calendar

My prompt:

What is the best time to plant Boro rice in Cachar district, and what variety should I choose?

Hermes Agent's r

esponse (summarised):

Nursery sowing: Mid-November to Mid-December
Transplanting: Mid-December to Mid-January
Harvest: April to May (before monsoon)

Recommended varieties specific to Assam: Joymati, Bishnu Prasad, Jyotiprasad, Kanaklata, IR-36, IR-50 — with maturity timelines and cold-tolerance tips for Silchar winters.

Sources for seeds: Assam Seeds Corporation, local block agriculture offices, KVK Cachar.

Demo 3: 3-Month Farming Plan

My prompt:

Create a simple 3-month farming plan for a rice farmer in Silchar starting from December

Hermes Agent's response:

Hermes generated a complete week-by-week plan across December, January, and February — covering seedbed preparation, cold protection, transplanting spacing (20x15cm), water management, fertilizer application timing, weed control, and pest monitoring. It even flagged the pro tip of draining water before Urea application to prevent runoff.

This is multi-step planning — not a single lookup but a structured, sequenced agricultural roadmap.

Why This Matters

Hermes Agent ran these three demonstrations on a phone in Termux. The same phone a farmer's son in Cachar district carries. The same environment that costs nothing beyond internet access.

The agentic capabilities that made this work:

Tool chaining — Hermes reasoned across multiple considerations (climate, soil, disease vectors, local institutions) in a single response without being explicitly prompted to do so.

Regional grounding — It applied Assam-specific knowledge (KVK Cachar, Assam Seeds Corporation, local variety names) rather than giving generic pan-India advice.

Honest uncertainty — It told the farmer to get a physical leaf sample confirmed before buying chemicals. An agent that knows the limits of its own remote diagnosis is more trustworthy than one that does not.

Planning capability — The 3-month farming plan was a genuinely structured multi-step output, not a list of generic tips.

Connection to My Broader Work

I have spent the past year building the Divine Earthly ASI system — a sovereign, offline-first agricultural AI for rural Indian farmers, running on quantized 0.5B models via llama.cpp on ARM64 Termux. My system is designed for zero-connectivity environments.

Hermes Agent occupies a different but complementary space: it requires internet connectivity but brings mature agentic infrastructure (skill learning, session memory, 28 tools) that would take years to build from scratch.

What they share: the belief that capable AI should run where farmers are, not just where servers are.

What Hermes Agent Does That Impressed Me

The skill-learning loop. After complex sessions, Hermes distills experience into reusable skill files. For agricultural use, this means the more farmers use it, the better it gets at local, regional questions — accumulating what I call Samskara: the impression that evolves future behavior.

See my companion Write submission for a deeper exploration of this Vedic parallel:
Samskara and the Self-Improving Agent

Reproduce This Yourself

Requirements: Android phone with Termux, or any Linux machine.

# Install proot-distro ubuntu (Termux only)
pkg install proot-distro
proot-distro install ubuntu
proot-distro login ubuntu

# Install Hermes Agent
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

# Set up free model (Google Gemini OAuth)
hermes model
# Select: Google Gemini via OAuth + Code Assist (free tier)

# Start
hermes

Then ask it a farming question. It works.

hermesagentchallenge #devchallenge #agents #ai #india

Samskara and the Self-Improving Agent: A Vedic Lens on Hermes Agent

Joydeep Das — Sat, 16 May 2026 14:51:38 +0000

submitted for the #hermesagentchallenge Write track

Where I Am Writing This From

I am writing this from Silchar, Assam, in Northeast India — on an Android phone, in Termux, with no laptop, no GPU, and no office.

I build AI systems for farmers. Farmers who ask questions like: Will it rain before my paddy is ready? Which disease is eating my leaves? Where can I sell my harvest today? These are not questions a ChatGPT subscription answers. These are questions that need an agent — one that runs offline, costs nothing to idle, remembers across sessions, and gets smarter the more it is used.

When I first read about Hermes Agent, I did not think about benchmarks. I thought about a Sanskrit word: Samskara.

What Is Samskara?

In Vedic philosophy, Samskara (संस्कार) refers to the impressions left on consciousness by experience. Every action leaves a trace. Every trace shapes future action. Over time, these accumulated impressions do not just record history — they evolve the entity itself.

This is not metaphor. In the Yoga Sutras of Patanjali, Samskaras are described as the actual mechanism by which a mind grows. You do not simply remember what happened. You become different because of what happened.

I have spent the past year building this concept into my AI architecture — the Divine Earthly ASI system — as a Samskara evolution cycle: a layer where the agent's experience actively reshapes its future reasoning patterns.

When I read Hermes Agent's core description, I almost dropped my phone.

Hermes Agent's Learning Loop Is Samskara

Here is what Hermes Agent does architecturally:

After a complex task involving five or more tool calls, a background process summarises the entire trajectory into a reusable Markdown skill file. This skill is stored on disk — readable, editable, inspectable. During future tasks, if that skill is found to be outdated or incomplete, it is patched in real time. A background curator then periodically reviews the entire skill library, consolidates overlapping skills, and archives stale ones.

Experience → Impression → Evolved Behavior.

That is Samskara. Not as decoration. As architecture.

The Western AI world calls this "self-improvement." The Vedic world described the same mechanism two thousand years ago as the primary engine of consciousness evolution. What Nous Research has built is not a new idea — it is an ancient one, finally implemented in code.

Evaluating Hermes Through the Five Pramanas

In Nyaya philosophy — the Indian school of logic — knowledge is validated through five Pramanas (sources of valid knowledge). I use this framework in my own benchmark work. Let me apply it to Hermes Agent.

Pratyaksha (Direct Perception)
Can the agent directly perceive its environment through tools? Yes. Hermes ships with 40+ built-in tools and MCP server integration, grounding its reasoning in real, observable data — not hallucinated assumptions.

Anumana (Inference)
Can it reason from evidence to conclusions across multiple steps? Yes. Multi-step planning and tool chaining are first-class capabilities, not afterthoughts bolted onto a chat interface.

Upamana (Analogy)
Can it recognise that a new problem resembles one it has solved before? Yes — this is precisely what the skill retrieval system does. Past solutions become templates for new situations.

Shabda (Testimony / Received Knowledge)
Can it learn from what others have told it, across sessions? Yes. Full-text search across all past conversations, plus a persistent user model that builds over time, means Hermes does not forget what you have taught it.

Anupalabdhi (Non-perception / Knowing what is absent)
Can it recognise the limits of its own knowledge? This is the hardest one. The skill curator's ability to archive stale knowledge is a step in this direction — knowing that something once learned is no longer valid is a form of epistemic humility most agents lack entirely.

Hermes Agent passes four of the five Pramanas cleanly. The fifth — genuine epistemic humility about the boundaries of its own knowledge — remains an open research frontier for all AI systems, including this one.

What This Means for a Farmer in Assam

Let me make this concrete.

Imagine a farmer asks an agent: My rice leaves have brown spots near the edge. What is wrong?

A static agent answers from its training data. It may be right. It may be wrong. Either way, it gives the same answer next week that it gives today.

A Samskara agent — a Hermes-style agent — does something different. It answers. Then it stores what it learned from this exchange. Next time a farmer in the same region asks a similar question during the same season, the agent does not start from zero. It starts from accumulated wisdom. The more farmers use it, the better it becomes for all farmers.

This is not a feature. This is a philosophy. It is the philosophy of the village elder who has seen fifty monsoons and knows things no textbook contains. Hermes Agent is, architecturally, a village elder that never forgets and never stops learning.

And here is the detail that made this personal for me: Hermes Agent installs on Android via Termux — the same environment I use every day to build and run AI systems. No laptop required. The same curl command that works on Linux works on my phone. This is not a tool built for Silicon Valley. This is a tool that can run where farmers actually are.

The Honest Comparison: What My Architecture Does Differently

I want to be honest here, not promotional.

My Divine Earthly ASI system runs fully offline on a quantized 0.5B parameter model using llama.cpp on ARM64 — no API calls, no internet required. Hermes Agent, in contrast, requires an LLM provider API key and internet connectivity for full functionality. For deep rural India where connectivity is unreliable, this is a real constraint.

What Hermes does better: the skill curation system is mature, inspectable, and battle-tested. My Samskara cycle is philosophically grounded but still experimental. Hermes has 134,000+ GitHub stars and a growing community. My projects have a growing vision and a long road ahead.

What I am building toward — sovereign, offline, Vedic-grounded agents for rural communities — is not yet fully achievable with Hermes Agent out of the box. But the architectural philosophy is the same. And that matters. When two systems built on opposite sides of the world, from different traditions, converge on the same core insight about how intelligent systems should learn — that convergence is worth paying attention to.

The Future I Am Building Toward

The agents that will matter most in the next decade are not the ones that answer questions. They are the ones that carry memory — not just of users, but of communities.

A Samskara agent for Indian farmers would carry the accumulated wisdom of ten thousand growing seasons. It would know which variety of paddy survives flooding in the Brahmaputra valley. It would know that the market price in Silchar drops three days after harvest in Cachar district. It would know these things not because a dataset told it, but because it learned them from real exchanges with real farmers, and those impressions deepened over time.

Hermes Agent's architecture points toward this. The skill system is the seed. The learning loop is the mechanism. The open-source, self-hostable design is the prerequisite for sovereignty.

We are not there yet. But the direction is right.

Getting Started

If you want to explore Hermes Agent:

GitHub: github.com/NousResearch/hermes-agent
Documentation: hermes-agent.nousresearch.com/docs
Android/Termux users: the installer auto-detects Termux — same curl one-liner as Linux

If you are building AI for communities that cannot afford to be forgotten by the next model update, Hermes Agent is worth your attention.

Joydeep Das is an independent AI researcher building sovereign, offline-first AI systems for Indian farmers under the Divine Earthly project. All development happens on an Android phone in Termux.

GitHub: github.com/divineearthly

hermesagentchallenge #devchallenge #agents #ai #india

I Built GPU-Free AI on a 3.4GB Android Phone

Joydeep Das — Sun, 10 May 2026 11:37:19 +0000

I built VedaRta entirely on a 3.4GB Android phone with Termux. No GPU. No cloud.

6 novel Vedic mathematical algorithms replace GPU-requiring operations:

Sphota Attention — 1,308× faster than O(n²) Softmax
Urdhva Matmul — 10.2× faster than BLAS on ARM64
Tri-Nadi Activation — Converges where SiLU explodes (loss 0.12 vs ∞)
Shunyam Norm — Zero-centered, no DC drift
Chitta KV Cache — 80% memory reduction
Katapayadi Encoder — Phoneme to vector

VedaRta Sphota is O(n) linear approximate attention — trades cross-token interaction for mobile efficiency. Different operation, different trade-off. Honest science matters.

"Aham Brahmasmi" produces PHI (1.6188) resonance from embeddings.
Trained a 49KB specialist model in 43 seconds on the phone.

GitHub: github.com/divineearthly/VedaRta
Model: huggingface.co/divinesouljoy/VedaRta-0.5B

I'm here to answer questions.