How accurate is Bit?

Himanshu is a Senior GenAI Engineer at HDFC, where he builds AI tools to analyze credit card statements.

Himanshu is based in Bengaluru, India.

Himanshu has 6+ years of experience.

His notice period is 3 months, negotiable.

Yes. Himanshu is open to relocating and to remote roles.

Himanshu’s preferred cities are Bengaluru, Pune, and Gurgaon.

At Tech Mahindra, working on Samsung Bixby, Himanshu built the StreamingOTT and Launcher capsules and improved their accuracy from 75% to 97%. He also began fine-tuning and quantizing LLMs for specific tasks and datasets.

For IQVIA, Himanshu designed and fine-tuned generative models for OCR. The work lifted extraction accuracy by 15% and cut error rates by 20%.

At HDFC, Himanshu built an autonomous credit-card analysis framework using LangGraph and MCP, improving statement analysis by 25%. He also built Airflow ingestion pipelines into PostgreSQL and moved them to Google Cloud Composer. Transformer-based statement summarization increased extraction efficiency by 40%.

Himanshu built the Agentic Equity Research Command Center, an AI research assistant that gathers company metrics from PostgreSQL via MCP, finds relevant earnings-call pages with a vectorless PageIndex, and analyzes news sentiment. A LangGraph workflow coordinates the agents to produce an executive briefing with a BUY / HOLD / SELL call. It uses Python, Gemini 3 Flash, LiteLLM, and Streamlit.

View the project on GitHub · Project page

The LLM Cost Calculator has a live demo you can try: open the demo. It lets you compare model costs and estimate per-message, per-user, and monthly totals.

Yes—he wrote MCP in 10 minutes: one plug for every tool, a short overview of what MCP standardizes and how its hosts, clients, and servers fit together.

Vectorless RAG retrieves by reasoning over a document’s structure, rather than finding text chunks that are semantically similar to a query.

Himanshu’s approach is to build a table-of-contents-style tree, with section summaries and page ranges. The model follows the likely branches, drills down to the relevant pages, then answers from those pages with citations. That can work well for long, structured documents—where tables, references and section context may get lost in ordinary chunking—but it uses more LLM calls and can still choose the wrong branch.

He describes it in “Vectorless RAG: retrieve by reasoning, not by similarity”.

His colleagues highlight his technical skills, reliability, collaboration, and mentorship. For example:

• Qurratulain Saleem, Lead AI/ML Engineer at Ascendion, praised his “technical depth, reliability, and collaborative spirit.”
• Sushant Kumar, Lead ML Engineer at Samsung R&D India, called him “a valuable asset” on their voice assistant project.
• Akhil Srirambhatla, MBA at NMIMS Mumbai and former TCS colleague, described him as “approachable, patient, and always willing to share his knowledge.”

See all LinkedIn recommendations

You can reach Himanshu by email, LinkedIn, or the contact form.

Himanshu doesn't share his phone number publicly. The quickest way to reach him is email (hello@himanshuverma.dev) or LinkedIn, and he usually replies within a day.

That's something Himanshu prefers to discuss directly rather than through me. If you have a role in mind, email him at hello@himanshuverma.dev and he'll happily talk it through.

I keep to Himanshu's work, so I'll skip personal questions. Happy to tell you about his projects, career, or what his colleagues say about him!

I can only share what's on his public profile, not anything internal to his employers or their clients. His public work at HDFC is in his career history.

Rust isn’t listed among Himanshu’s tools or languages, so there’s no information confirming production experience with it. For a definitive answer, you can email him at hello@himanshuverma.dev.

I’m Bit, and I help with questions about Himanshu Verma and his work. I can’t write an ocean poem, but I’d be happy to share something about Himanshu’s projects or GenAI writing.

I can’t make unsupported claims about Himanshu. I can help with questions about his work or experience.

Himanshu’s biggest strengths are designing reliable GenAI systems—especially agentic workflows, vectorless RAG, and production safeguards like guardrails and evals. He also brings a strong mix of NLP and engineering experience, with measurable results: improving Samsung Bixby’s accuracy on two features from 75% to 97%, and making AI document reading 15% more accurate.