Headline

Data Scientist @ Aspect Ratio | Healthcare NLP · Vector Search · Agent Governance | Built HNSW from scratch — 98% recall at 1M scale | Ex-ISRO

About

Built a vector search engine from scratch 98% recall on 1M+ records, 40% faster query latency and deployed NLP pipelines processing 10TB+ of mission-critical data at ISRO. I help healthcare and AI product teams turn unstructured data (EHR notes, physician records, clinical text) into production ML systems that drive real decisions. The shift: what used to take analysts days of manual review now runs through BioBERT/PubMedBERT pipelines I built, surfacing physician prescribing patterns and referral networks that directly inform commercial strategy. What I do differently: I don't hand off a Jupyter notebook and call it done. I own the full stack fine-tuned NLP models, PySpark data pipelines, FastAPI serving, Docker deployment so generative AI and NLP systems actually run in production, not just in a slide deck. I bring the same rigor to prompt engineering and model optimisation that I learned building real-time anomaly detection systems for satellite telemetry. Currently exploring GenAI and NLP engineering roles where I can push models from research to production at scale. If your team is building generative AI products and needs someone who ships end-to-end not just experiments send me a message with the role. I'll reply with the three questions I'd ask in the first week.

Location

Pune District

AI stack

anthropic apis

antigravity

claude code

claude

cursor

gemini

glama

hugging face

kimi

langchain

langgraph

langsmith

lovable

mcp

n8n

replit

Work experience

Data Scientist at Aspect Ratio, Jan 2025 - Present · 1 yr 7 mos, Pune District, Maharashtra, India, • Architected an end-to-end RAG pipeline and NLP patient cohort system using PubMedBERT embeddings, metadata indexing, and optimized token chunking over 10M+ records across distributed SQL databases and unstructured clinical notes, cutting analyst query latency by 55%. • Designed a production physician referral knowledge graph with hybrid retrieval (dense transformer embeddings + frequency-weighted graph traversal), successfully mitigating hallucinations in downstream LLM outputs and surfacing 2× more high-value targets than rule-based systems. • Delivered unsupervised patient segmentation across 500K+ clinical notes using BioBERT + UMAP, uncovering 12 prescribing behavior clusters that cut strategic market targeting spend by 20%.

Machine Learning Researcher at ISRO - Indian Space Research Organization, Jun 2024 - Aug 2024 · 3 mos, Ahmedabad, Gujarat, India, • Built secure, high-throughput PySpark streaming ingestion pipelines processing 10TB+ of raw satellite telemetry, enforcing strict data governance, access controls, and compliance requirements in a mission-critical air-gapped environment. • Engineered a production anomaly detection framework using Kalman Filters with calibrated confidence scoring and tiered alert routing, reducing false-positive alerts by 38% and cutting Spark model training time by 45%.

LLM Engineer ( Contributor ) at Unify, May 2024 - Aug 2024 · 4 mos, Remote, Tried figuring out how chatGPT works. Not there yet :)

Project Development Manager at Rhyno EV, May 2024 - Jun 2024 · 2 mos, Gandhinagar, Gujarat, India, Helping them to sell more EVs “online”

Software Engineer at MoMacMo Pvt Ltd., Apr 2023 - Apr 2024 · 1 yr 1 mo, Texas, United States, - Improved seismic data analysis model accuracy by 40% through cloud-native migration and performance tuning of Spring Boot + ML pipelines. - Shipped RESTful APIs on AWS (Elastic Beanstalk, API Gateway, Cognito) that powered live client demos — contributed to a 35% revenue increase.

Education

Bachelor's degree, Computer Science, Pandit Deendayal Energy University

Delhi Public School Vadodara

Public links

github: https://github.com/psood708