Senior AI/ML Engineer with 4+ years of experience in Generative AI & LLM Systems. S.D. is experienced in leading ML training platforms, developing graph-based multi-agent systems, automating legal-document drafting, and creating multi-page legal archive digitization & reconstruction systems. S.D. has boosted patent attorney productivity and developed conversational AI & NER tooling.
Led a team of 8 Python developers to develop distributed ML pipelines and a custom GPU VRAM partitioning algorithm, enabling an estimated initial cost saving of 80% using consumer-grade GPUs.
Automated 30% of HR operations by designing leave + onboarding Agentic AI modules for the internal "Second Brain" platform, incorporating tool-calling, memory-based reasoning, multi-step planning, and fallback guards.
Delivered a $400K+ annual OCR savings by designing an in-house GPU-accelerated OCR + layout extraction system deployed on AWS ECR, replacing costly Textract usage.
Designed leave + onboarding Agentic AI modules for our internal “Second Brain” platform using LangGraph, incorporating tool-calling, memory-based reasoning, multi-step planning, fallback guards, and secure departmental vector “brains,” demonstrating end-to-end ownership of agentic AI system design that powers enterprise knowledge automation.
Slashed drafting time by 60%: Built an LLM-powered drafting system enabling NZ lawyers to
pay taxes and generate 31 categories of legal notices, claims & affidavits; designed modular
prompt-templating workflows and selected O4 Mini, reducing hallucinations by 45% and
improving legal-text reliability with citation-enriched outputs.
Boosted patent attorney productivity by 30%: As early adopters of first-generation LLMs, led
QLoRA finetuning of LLaMA, Falcon, Mistral & Mixtral; extended context using
YARN/RoPE improving long-claim coherence by 25%, and built an LLM-based SVG generator
achieving 94% diagram accuracy.
Delivered 99.2% intent accuracy across 13 departments: Contributed to the enterprise-wide
CossBot assistant, implementing a multilingual fuzzy spell-check (English, German, French),
scalable NER for 2000+ entities, and a transformer-based classification pipeline achieving 93%
extraction accuracy on real-world queries.
Education
M.Tech (Computer Science)
IIT Gandhinagar
2020 - 2022 (2 years)
B.Tech (Computer Science)
TMSL Kolkata
2016 - 2020 (4 years)
Image Caption Generator using Siamese Graph Convolutional Networks and LSTM
