DEV Community: Navchetna Technologies

How Osmium AI Works: Building an Education-First LLM Instead of Another Chatbot

Navchetna Technologies — Sun, 26 Jul 2026 20:42:41 +0000

At the heart of OsmiumLLM is a Dynamic Mixture of Experts (MoE) architecture.

Unlike traditional dense models that activate every parameter for every request, OsmiumLLM intelligently activates only the experts required for the task.

Think of it like a hospital. When you walk in, you don't meet every doctor. You meet the specialist who understands your problem.

OsmiumLLM works the same way.

Step 1: Intelligent Routing
Every request first passes through an intelligent routing layer. It evaluates:

subject domain

reasoning complexity

context length

educational intent

compute availability

The routing engine then decides which expert networks should solve the problem. Simple questions use lightweight experts. Complex reasoning activates additional specialists. This keeps latency low while maintaining high-quality responses.

Step 2: Dynamic Expert Selection
Instead of using all 96 billion parameters, OsmiumLLM dynamically activates only the subset required for the request.

For example:

factual question → lightweight experts

mathematics → reasoning experts

programming → code experts

educational explanation → tutoring experts

long-context analysis → extended-context experts

This architecture improves both efficiency and scalability.

Step 3: Memory Optimization
Inference isn't just about generating tokens. It's also about using hardware efficiently.

OsmiumLLM dynamically optimizes:

KV cache allocation

memory consumption

context utilization

expert activation

inference scheduling

This allows longer conversations while minimizing unnecessary compute.

The Osmium AI Inference Platform
OsmiumLLM is served through Osmium AI, our production inference platform.

Rather than exposing a single generic model, developers can choose the right model for their workload.

Chat Models (organised by use‑case)
Osmium Fast – Ultra‑fast applications • 131K context window • Ideal for chatbots, real‑time assistants, and high‑throughput APIs.

Osmium Lite – Cost‑efficient deployments • 32K context window • Designed for lightweight applications where affordability is the priority.

Osmium Balanced – General production workloads • 131K context window • Provides an optimal balance between speed, capability, and cost.

Osmium Code – Programming & software engineering • 262K context window • Optimised for code generation, debugging, refactoring, and technical documentation.

Osmium Full – Advanced educational reasoning • 131K context window • Built for complex academic explanations, concept breakdowns, and long‑form educational assistance.

Osmium Agent – AI agents & tool calling • 131K context window • Supports function calling, multi‑step workflows, and autonomous agent applications.

Osmium Reasoning – Deep reasoning & complex problem solving • 131K context window • Our highest‑capability model, designed for advanced reasoning, research, and multi‑step analytical tasks.

Embedding Models
Osmium Embed – General‑purpose text embeddings • 8K context window

Osmium Embed Code – Code search & software intelligence • 8K context window

Complete model metadata is publicly available through our Models API.

Built for Developers
We wanted developers to integrate Osmium AI in minutes—not weeks.

That's why the platform is fully OpenAI API compatible.

Supported Features:
✅ Streaming responses

✅ Function calling

✅ JSON mode

✅ Long‑context inference

✅ Embeddings

✅ Moderation

✅ Intelligent fallback routing

✅ Request tracing

Existing OpenAI SDKs work with only a base URL change.

Intelligent Model Fallback
Reliability matters.

If a preferred model becomes temporarily unavailable, Osmium AI automatically routes requests to the next most appropriate model.

For example:

text
osmium-reasoning → osmium-balanced
without requiring any changes from the application.

The result is higher availability and fewer production failures.

Infrastructure
Our inference platform is powered by a custom gateway built in Rust (Axum).

The architecture includes:

OpenAI‑compatible API layer

intelligent routing engine

automatic model failover

secure authentication

streaming infrastructure

request tracing

multilingual processing

The infrastructure is designed to deliver consistent low‑latency inference while remaining scalable for enterprise deployments.

Built for Indian Education
One thing we observed while building OsmiumLLM was simple.

Indian students don't always ask questions the way benchmark datasets expect. They switch between Hindi and English. They write formulas without context. They use abbreviations. They mix concepts from different chapters.

General‑purpose LLMs often struggle here.

OsmiumLLM is optimised for multilingual educational workloads and natural student interactions, making it particularly effective for Indian classrooms and competitive exam preparation.

Beyond Chatbots
Osmium AI isn't limited to student conversations.

Organizations are using the platform for:

AI tutors

Educational assistants

Learning Management Systems

Universities

Assessment platforms

Developer tools

Enterprise knowledge systems

Coding assistants

Institutional intelligence

Built for Enterprises
For organizations building AI‑powered educational products, Osmium AI offers:

OpenAI‑compatible APIs

Dedicated inference capacity

Private deployments

Enterprise SLAs

Custom rate limits

Volume pricing

Priority support

Whether you're building an EdTech platform, an internal learning system, or integrating AI into existing products, Osmium AI is designed to scale with enterprise requirements.

For partnerships or enterprise deployments, contact: hello@navchetna.tech

Learn More
If you'd like to dive deeper into OsmiumLLM's architecture, educational philosophy, and technical design, explore our interactive knowledge page:

🔗 https://osmium.co.in/share/mruysi88br4p3

Browse the complete model catalog and API documentation:

📚 Models: https://api.navchetna.tech/v1/osmium/model-list

🚀 Try Osmium AI: https://osmium.co.in

The Bigger Vision
We don't believe AI should replace education. We believe it should make great education accessible to everyone.

Every explanation should be clearer. Every doubt should get an answer. Every student should have access to guidance whenever they need it.

That's the vision behind Osmium AI and OsmiumLLM.

And we're only getting started.

Explore Osmium AI: https://osmium.co.in

Enterprise & Partnerships: hello@navchetna.tech

We Didn't Build Another AI Chatbot. We Built an AI Study Companion for Future Doctors.

Navchetna Technologies — Sun, 26 Jul 2026 20:14:31 +0000

"Every doctor once sat exactly where today's NEET aspirants are sitting—surrounded by books, uncertainty, and an impossible dream."

Every year, millions of students prepare for NEET.

Most of them aren't chasing marks.

They're chasing a white coat.

They're chasing the moment their parents hear, "Congratulations, you've been selected."

But getting there isn't easy.

The Hardest Part of NEET Isn't Biology

Ask any NEET aspirant.

The challenge isn't finding study material.

It's staying consistent for months.

Thousands of pages.

Hundreds of diagrams.

Endless mock tests.

Revision after revision.

Eventually every student reaches a point where they stop asking,

"Can I crack NEET?"

and start asking,

"Am I even good enough?"

That's the moment where most students quietly give up.

AI Can Answer Questions.

But Can It Help Someone Learn?

When generative AI exploded, education platforms rushed to integrate chatbots.

We did too.

Students loved getting instant answers.

But after analyzing how they used the platform, we noticed something surprising.

Students kept asking the same questions.

Making the same mistakes.

Repeating the same misconceptions.

The AI answered perfectly.

Yet learning wasn't improving.

That's when we realized something.

Education isn't about generating answers. It's about building understanding.

Building an AI That Thinks Like a Mentor

Instead of building another chatbot, we asked ourselves a different question.

What if AI understood why a student was confused before trying to explain the answer?

That philosophy became the foundation of Osmium AI.

Instead of immediately responding, the platform focuses on understanding:

where the conceptual gap exists
why the student made the mistake
whether they need an analogy, diagram, or deeper explanation
what topic they should revise next

Learning isn't linear.

AI shouldn't be either.

Engineering Behind the Experience

Since this is dev.to, here's what powers the platform.

Osmium AI runs on a custom inference platform built with Rust and Axum, exposing OpenAI-compatible APIs for developers.

Highlights include:

Streaming inference
Function calling
JSON mode
Intelligent model routing
Automatic model failover
Multilingual support
Specialized reasoning, coding, embedding, and moderation models

Rather than treating every request equally, workloads are routed to models optimized for their specific purpose.

Built for Indian Students

Students preparing for NEET rarely ask textbook-perfect questions.

Sometimes they type half a Biology diagram.

Sometimes they switch between English and Hindi.

Sometimes they only remember part of a question.

Educational AI has to understand students before answering them.

That's why multilingual learning and educational reasoning became core design goals instead of optional features.

AI Won't Replace Great Teachers

Teachers inspire.

Mentors motivate.

Parents sacrifice.

AI can't replace any of them.

But AI can make sure a student never has to wait until tomorrow to understand today's doubt.

That's the future we believe in.

For Every Future Doctor

If you're preparing for NEET, we've built a dedicated AI-powered experience with concept explanations, doubt solving, revision support, mock preparation, and personalized learning guidance.

Explore it here: https://osmium.co.in/for/neet

Because sometimes one concept understood today becomes one life saved tomorrow.

I Asked 100 JEE Aspirants One Question. Their Answers Changed How We Built AI.

Navchetna Technologies — Sun, 26 Jul 2026 20:01:31 +0000

You don't fail JEE because you don't study. You fail because, somewhere along the way, you stop believing you can crack it.
That sentence came from one of the students we spoke to.
It stayed with us.
Over the last few months, we spoke with more than 100 JEE aspirants - from coaching institutes, online communities, and students preparing alone from home.
We expected to hear about difficult subjects.
Instead, we kept hearing something else.
Students weren't struggling because they lacked information.
They were struggling because they felt lost.
If you've ever prepared for JEE, this story is for you.
Not as a founder.
Not as a developer.
Just as another student trying to make sense of one of India's toughest exams.

Imagine This
It's 11:47 PM.
Tomorrow's coaching class starts at 8:00 AM.
You've spent the last 40 minutes staring at the same Organic Chemistry question.
Nothing clicks.
You open YouTube.
The explanation is 52 minutes long.
You search Google.
Five websites explain it five different ways.
You ask ChatGPT.
It gives you the correct answer instantly.
But somehow…
you still don't understand why your own approach was wrong.
So you close your laptop.
Tell yourself you'll revise it tomorrow.
Tomorrow becomes next week.
That one weak concept quietly survives every revision session until it finally costs you 12 marks in the exam.
That isn't laziness.
That's how confidence disappears - one unresolved doubt at a time.

We Thought AI Would Solve This
Like almost everyone entering the AI space, we started with a chatbot.
It could:
Solve numerical problems
Explain concepts
Answer theoretical questions
Generate step-by-step solutions

Students loved it.
For about a week.
Then something unexpected happened.
The same students returned.
With the same doubts.
The same misconceptions.
The same mistakes.
That forced us to ask a difficult question.
If students already have answers, why aren't they learning?
Eventually we realized something uncomfortable.
Answering questions doesn't automatically create understanding.

The Problem Isn't Intelligence
It's Isolation.
JEE is one of the loneliest journeys a student can experience.
Millions prepare every year.
Yet almost everyone feels like they're fighting the battle alone.
When students get stuck, they usually don't need another PDF.
They don't need another playlist.
They don't even need another answer.
They need someone who can understand where their thinking broke.
That's what great teachers do.
They don't simply tell you the answer.
They identify the misconception hiding underneath the question.
We wanted AI to do the same.

So We Stopped Building Another Chatbot
Instead, we started asking a different question.
What if AI tried to understand the student's mistake before solving it?
That single question changed almost everything.
Instead of immediately producing solutions, our system tries to understand:
Which concept is actually missing?
Is this a conceptual mistake or just a calculation error?
Has the student made this mistake before?
What's the shortest path to fixing this permanently?
What explanation is most likely to be remembered during the exam?

Because every wrong answer tells a story.
Most AI systems ignore that story.

Meet Osmium AI
Osmium AI wasn't built to impress users with long responses.
It was built around a much simpler objective:
Help students remember concepts when they actually need them - in the examination hall.
Whether you're studying:
Mechanics
Electrostatics
Organic Chemistry
Coordinate Geometry
Modern Physics
Calculus

the goal stays exactly the same.
Not just answering.
Teaching.

Built Around How Indian Students Actually Learn
Students rarely ask perfect questions.
Sometimes they upload blurry notebook images.
Sometimes they write only half an equation.
Sometimes they mix Hindi and English.
Sometimes they remember the chapter but not the concept.
Sometimes they don't even know what they're supposed to ask.
That's normal.
Instead of expecting perfect prompts, we designed the experience around imperfect ones.
Because that's how real students learn.

The Engineering Behind It
For the developers reading this, here's a quick overview of the platform powering the experience.
Custom inference platform built in Rust (Axum)
OpenAI-compatible APIs
Intelligent model routing
Streaming responses
Function calling
JSON mode
Automatic failover
Dedicated reasoning models
Coding models
Embedding models
Moderation models
Optimized multilingual inference for Indian education

None of those technologies exist for marketing slides.
They exist for one reason.
When a student is stuck on a problem at midnight, waiting 15 seconds feels much longer than it sounds.
Learning should feel immediate.

Technology Isn't the Goal
Education is.
We're not trying to replace teachers.
Teachers inspire.
Teachers mentor.
Teachers understand students in ways software never fully can.
Our goal is much smaller.
If AI can help one student understand Electrostatics today…
that student might enter the examination hall six months later with confidence instead of fear.
Sometimes a single concept understood at the right time changes everything.

If You're Preparing for JEE
We've built a dedicated AI experience specifically for engineering aspirants.
It focuses on:
Personalized concept explanations
AI-powered doubt solving
Mistake analysis
Guided learning
Exam-oriented understanding

Instead of giving faster answers, we want to help students build stronger understanding.
Because the difference between clearing JEE and missing it often isn't intelligence.
It's confidence.
And confidence usually begins with finally understanding that one concept you've been avoiding.

If you'd like to explore what we've been building, you can learn more at https://osmium.co.in/for/jee.
We'd genuinely love to hear your thoughts, feedback, and experiences as a student or educator.
After all, this journey started by listening to students.