Gitika Bhatia

Full Stack Developer
Frontend + Backend ⚡
Always learning

[Gitika Bhatia @gitika_dev]

Full Stack Developer with 3+ years of experience, building and shipping production-ready web applications end-to-end.

Sharing my portfolio👇
🔗https://t.co/DsqxCQaqs8

#FullStackDeveloper #BuildInPublic

[Gitika Bhatia @gitika_dev]

I've got hardcoded checks for known breakers — like "Create an invoice" with zero details, which used to make the AI invent a fake one. That goes straight to chat now.
Everything else goes to a small AI model with examples of what's a command vs just talking. "What can you do?" is actually one of those examples — no invoice keywords, so it needed an explicit rule or the model would guess wrong.
Pure AI tripped on vague stuff, pure rules can't cover every phrasing - so I landed on both.

[Gitika Bhatia @gitika_dev]

Update on Ledger 👇

One problem with the invoice agent: not every user message is an invoice command.

"Create an invoice" (without details) could trigger bad assumptions.
"What can you do?" had nowhere to go.
GST/VAT questions weren't handled well.

So I added a Chat Node.
✅ Separates conversation from invoice actions
✅ Answers capability & tax questions naturally
✅ Asks clarifying questions for incomplete requests
✅ Prevents fake invoice generation

The bigger improvement was upstream: adding deterministic intent checks before the LLM sees the request. Fewer guesses, better routing, safer outcomes.

Small feature. Big reliability upgrade.
#buildinpublic #AI #LangGraph #LangChain

[Gitika Bhatia @gitika_dev]

The full architecture of Ledger — an AI invoicing agent.
Auth → onboarding → 6-node LangGraph pipeline → invoice detail
Invoice state is persisted: draft → confirmed → sent → paid
Live at https://t.co/mLSGVNC8tI

[Gitika Bhatia @gitika_dev]

Built Ledger — an AI invoicing agent. Just type what you need.
Under the hood:
→ LangGraph pipeline
→ Router → Generator → Editor → RAG
→ Multi-invoice support & Taxes back calculation
→ Vector search across past invoices
→ INR, USD, EUR

More coming soon.
#buildinpublic

[Gitika Bhatia @gitika_dev]

Invoicing shouldn't take 30 minutes.

Built something about that.

The vision: describe your invoice in simple English — AI handles the rest.

Demo dropping soon 👀

#buildinpublic #AI

[Gitika Bhatia @gitika_dev]

I built a LangGraph agent with human approval today.

Here’s how it works:
1️⃣ User enters a task
2️⃣ Agent generates a step-by-step plan
3️⃣ Workflow pauses for approval

✅ Approve → Agent executes the plan
❌ Reject → User is asked to run the agent again

One thing I'm learning while building agents:
Generation is easy.
Control is the real challenge.

LangGraph makes it much easier to design workflows where agents can pause, wait for humans, and then continue execution.

#LangGraph #AIAgents #LangChain #BuildingInPublic

[Gitika Bhatia @gitika_dev]

LangGraph becomes much easier once you understand these 3 concepts:

1️⃣ State Shared memory across the workflow
2️⃣ Nodes Actions like LLM calls or tools
3️⃣ Edges Logic that decides the next step

Together they create stateful AI workflows with loops, retries, and branching.

#AI #LangGraph #AIAgents #LearningInPublic

[Gitika Bhatia @gitika_dev]

@ronniehuss Exactly. Chains handle pipelines well, but real agents need state and decision-making.

[Gitika Bhatia @gitika_dev]

Many AI applications start with simple LLM chains.

But real AI apps need more:
- multiple steps
- decision making
- retries
- loops
- state across interactions

LangChain simplifies building LLM apps with tools, memory, and structured pipelines using LCEL.

But when workflows get complex, LangGraph helps build stateful AI agents with branching, loops, and decision-making.

LCEL = pipelines
LangGraph = agent workflows

#AI #LangChain #LangGraph #AIAgents

[Gitika Bhatia @gitika_dev]

After learning about RAG, Chunking, and Embeddings, I built a small Light RAG demo to see how it works in practice.

Goal:
Make the model answer questions only from the documents we provide instead of relying purely on its training data.

How it works:

Ingestion Phase
- Add a document / policy text
- Split it into chunks
- Generate embeddings
- Store them in a vector store

Query Phase
- User asks a question
- Convert query → embedding
- Retrieve similar chunks via semantic search
- Send retrieved context to the LLM → generate answer

Tech Stack:
- Next.js
- Node.js
- LangChain
- OpenAI Embeddings
- In-memory Vector Store

Building small demos like this is helping me understand how LLMs + external knowledge power modern AI applications.

#AI #RAG #LangChain #BuildInPublic #LearningInPublic

[Gitika Bhatia @gitika_dev]

How do AI systems search through massive documents and find the right answer?

Two key ideas make it possible:

Chunking
Large documents are split into smaller sections so the system can search specific pieces of information.

Embeddings
Each chunk is converted into a numerical vector that captures its meaning.

When a user asks a question, the system compares the query embedding with stored chunk embeddings to find the most relevant context.

This is the core mechanism behind RAG (Retrieval Augmented Generation) used in many modern AI applications.

#AI #RAG #Chunking #Embeddings #BuildInPublic #LearningInPublic

[Gitika Bhatia @gitika_dev]

@modelstoreai Thanks a lot, really appreciate it! 🙌
That sounds interesting — will check it out for future experiments 🚀

[Gitika Bhatia @gitika_dev]

Built my first full-stack AI project: Promptify 🚀

React + Node + MongoDB + OpenAI

Chat + image generation in one app.

From learning concepts to building real products — this feels good.

Learning in public. Building daily.
#BuildInPublic #AI

[Gitika Bhatia @gitika_dev]

@enrique_turman Thanks! 🙌
Completely agree — shipping real projects teaches the most. On to the next build 🚀

[Gitika Bhatia @gitika_dev]

@helderbuilds Totally agree — building things hands-on makes the concepts much clearer than just reading about them.

[Gitika Bhatia @gitika_dev]

Practiced OpenAI hands-on today 🚀

Built a basic AI chatbot from scratch to understand how LLMs actually work in an app.

Small project, big clarity.

#BuildInPublic #AI #OpenAI

[Gitika Bhatia @gitika_dev]

Ever wondered how AI tools answer questions using your own documents?
The answer is RAG
AI retrieves relevant info first, then generates the answer.
Docs → Chunking → Embeddings → Vector DB
Query → Similarity Search → Context → LLM → Answer
#AI #RAG #LLM #BuildInPublic

[Gitika Bhatia @gitika_dev]

Built a Search Web Agent 🤖🔎

Next.js + OpenAI + LangChain + Tavily.

LLM answers simple queries.
Automatically switches to web search when needed.

Trying to explore how AI agents can combine LLMs + real-time web data.

#AI #LangChain #BuildInPublic

[Gitika Bhatia @gitika_dev]

Building AI agents in production?

LangChain simplifies:

→ Prompt management
→ Tool calling
→ Memory handling
→ Streaming responses
→ Multi-step reasoning

With LCEL, complex workflows become clean, structured pipelines.

#LangChain #LLM #AIEngineering #JavaScript

[Gitika Bhatia @gitika_dev]

Spent time today building clarity around modern AI fundamentals.

LLMs, Generative AI, Agentic AI, AI agents, workflows, multi-agent systems, and RAG.

Not deep dives — but enough to understand how intelligent systems are structured beyond simple prompting.

#BuildInPublic #AI