Headline

Building Efficient AI Systems | Large Language Models • Distributed Training • GPU Optimization

About

I like to research and build stuff

Location

Bengaluru

AI stack

claude code

cursor

grok

hugging face

langchain

mcp

Work experience

Machine Learning Engineer at Nuroen.AI, Jul 2026 - Present · 3 mos, Bengaluru, Karnataka, India, Fine-tuned a 32B model on 8×A100 over a 2.05M-sample instruction corpus I built and decontaminated myself: +24 pts instruction-following, +16 grounded refusal, +4 tool-calling — and I reported the −8.5 MMLU-Pro regression alongside them, because an eval that only publishes gains is marketing. Built a paired A/B harness to test whether retrieving auto-generated agent tool-skills actually helps: 1,088 execution-verified tasks → 93.5% to 99.0% success (+5.5pp, McNemar p<0.001), then shipped a retrieval gate that held the entire win at 87% lower token cost. Validated that against the 50+ real production connectors (181 actions) — and the synthetic numbers didn't hold: retrieval found the right tool 100% of the time but ranked it first only 68%. The run caught a bug in my own shipped default, silently dropping the correct tool 18.8% of the time. Fixed before it reached production. Removed the scaling wall on a real-time STT→LLM→TTS voice agent: latency flat 1.3–1.8s from 1→128 concurrent (replacing a stack that ballooned 161ms → 3.3s), sustained soak of 240/240 turns at p95 215ms, and accent parity at 8.4% WER for Indian-accented English vs 7.4% for US. Migrated a 1.14M-chunk knowledge base to Postgres + pgvector: −49% storage, 4.4× faster, zero re-embedding — and proved it safe rather than assuming (top-1 92% parity vs exact search), which surfaced two real bugs including a tuning parameter that was silently a no-op. Made the vector store swappable and proved it: pgvector vs Qdrant recall identical to 4 decimal places. The commercially useful conclusion — retrieval quality is the embedder and chunking, not the vendor — so no lock-in.

Data Processing Intern at CHRIST University, Yeshwanthpur Campus (BYC), Jan 2026 - Mar 2026 · 2 mo, Bengaluru, Karnataka, India, Conducted research in LLMs focusing on bias evaluation and mitigation in low-resource indian languages

Summer Intern at Calcutta Mathematical Society, Jun 2023 - Aug 2023 · 3 mos, Kolkata, West Bengal, India

Education

Master of Science, Data Science, Christ University, Bangalore

Bachelor of Science, Mathematics, Amity University Kolkata

Public links

github: https://github.com/priangshu17