Built my first project in the Generative AI in Finance course: a voice-enabled Personal Finance AI Advisor
• RAG on SEBI and GST documents
• Live gold/silver prices in ₹
• Agent that picks docs, prices, or both
• Speak your question, hear the answer
Day 1 of Generative AI in Finance.
JPMorgan cut 360,000 hours/year of manual loan review with AI. Barclays blocks fraud before transactions even complete. Fraud takes milliseconds, and human review takes hours.
Now I’m building the same kind of systems.
Starting the Generative AI in Finance course today 🚀
Over the next few weeks I’ll be building a RAG finance advisor, an AI stock trading agent, and a fraud detection model fine-tuned with LoRA.
Excited to see what AI can do when the cost of being wrong is real money.
Day 2 of learning NLP: text preprocessing.
Mind-bending part: removing "stopwords" can accidentally flip a sentence's meaning.
"Not bad" → remove "not" → "bad"
Same words, opposite meaning. Machines don't get sarcasm, context, or nuance for free. You have to engineer it in.
Day 1 of learning ML.
Ended up going deep into NLP instead of just skimming fundamentals.
Wildest part: teaching a machine "it's raining cats and dogs" isn't rain or animals, it's just... rain. Idioms break these models more than actual grammar does.
Onwards. 🧵 more soon
Built an AI agent for India's new criminal code (BNS) using LangChain + RAG.
It decides on its own: search the BNS PDF for sections/punishments, search the web for recent judgments, or both — and remembers context across the conversation via checkpoint
🔗 https://t.co/NUFAW7jtO0
Built a RAG app with LangChain - load web docs, chunk them, embed with HuggingFace, store in ChromaDB, then query with Gemini.
The model only answers from retrieved context, not its own training data. Watching retrieval actually pull the right chunks made RAG finally click.