Looks like you have JavaScript disabled. For the full Twine experience, you will need to re-enable it.

I am Priyanshu Kumar, an AI Engineer with 5+ years of hands-on experience building and deploying production AI systems across financial research, insurance claims, and transaction risk analytics. I currently work at BlackRock on retrieval and summarization platforms, designing vector-based search, LLM-assisted query rewriting, and evaluation pipelines that measurably improve analyst productivity and answer quality. Previously I delivered NLP and ML solutions at AIG for claims intake automation, severity signals, and fraud enrichment, improving data capture accuracy and reducing manual rework. I am known for Python, applied machine learning, NLP, model evaluation, and production APIs, translating messy real-world data into reliable, decision-ready systems without over-engineering.…I am Priyanshu Kumar, an AI Engineer with 5+ years of hands-on experience building and deploying production AI systems across financial research, insurance claims, and transaction risk analytics. I currently work at BlackRock on retrieval and summarization platforms, designing vector-based search, LLM-assisted query rewriting, and evaluation pipelines that measurably improve analyst productivity and answer quality. Previously I delivered NLP and ML solutions at AIG for claims intake automation, severity signals, and fraud enrichment, improving data capture accuracy and reducing manual rework. I am known for Python, applied machine learning, NLP, model evaluation, and production APIs, translating messy real-world data into reliable, decision-ready systems without over-engineering.

Priyanshu Kumar

AI Engineer, Data Scientist, Full Stack Developer, +2





I am Priyanshu Kumar, an AI Engineer with 5+ years of hands-on experience building and deploying production AI systems across financial research, insurance claims, and transaction risk analytics. I currently work at BlackRock on retrieval and summarization platforms, designing vector-based search, LLM-assisted query rewriting, and evaluation pipelines that measurably improve analyst productivity and answer quality. Previously I delivered NLP and ML solutions at AIG for claims intake automation, severity signals, and fraud enrichment, improving data capture accuracy and reducing manual rework. I am known for Python, applied machine learning, NLP, model evaluation, and production APIs, translating messy real-world data into reliable, decision-ready systems without over-engineering.…I am Priyanshu Kumar, an AI Engineer with 5+ years of hands-on experience building and deploying production AI systems across financial research, insurance claims, and transaction risk analytics. I currently work at BlackRock on retrieval and summarization platforms, designing vector-based search, LLM-assisted query rewriting, and evaluation pipelines that measurably improve analyst productivity and answer quality. Previously I delivered NLP and ML solutions at AIG for claims intake automation, severity signals, and fraud enrichment, improving data capture accuracy and reducing manual rework. I am known for Python, applied machine learning, NLP, model evaluation, and production APIs, translating messy real-world data into reliable, decision-ready systems without over-engineering.

Available to hire

I am Priyanshu Kumar, an AI Engineer with 5+ years of hands-on experience building and deploying production AI systems across financial research, insurance claims, and transaction risk analytics. I currently work at BlackRock on retrieval and summarization platforms, designing vector-based search, LLM-assisted query rewriting, and evaluation pipelines that measurably improve analyst productivity and answer quality.

Previously I delivered NLP and ML solutions at AIG for claims intake automation, severity signals, and fraud enrichment, improving data capture accuracy and reducing manual rework. I am known for Python, applied machine learning, NLP, model evaluation, and production APIs, translating messy real-world data into reliable, decision-ready systems without over-engineering.

Skills

Experience Level

Expert

Expert

Expert

Expert

Expert

Expert

Expert

Expert

Expert

Expert

Expert

Language

English

Fluent

Work Experience

Applied AI Engineer at BlackRock

February 1, 2025 - Present

Built a document retrieval layer in Python using a Pinecone vector index to search filings, portfolio commentary, and research notes, improving Top-3 result relevance by 32% in internal evaluation runs. Enhanced search accuracy for ambiguous analyst queries via LLM-based query rewriting and BM25 hybrid reranking, increasing Top-1 match rate by 14% for short questions. Implemented structured coverage checks to enforce required tickers, risk terms, and date windows in summaries, boosting completeness from 71% to 88% across a 200-sample review. Established weekly regression testing on ~500 benchmark questions with RAGAS-style metrics and a custom financial rubric, blocking 3 prompt/template releases after detecting regressions. Reduced latency during market-open spikes by caching high-frequency tickers and briefs in Redis, cutting repeat-request latency by 45%.

Claims AI Engineer – Claims Intelligence & Automation Engine at AIG

September 1, 2024 - January 31, 2025

Built a claims-intake extraction pipeline using Python, spaCy, and Hugging Face models to convert adjuster notes into structured fields (incident type, injury severity, loss narrative), improving usable field capture from 62% to 87% across auto and property claims. Added a prompt-guided fallback flow using LLM APIs for poorly formatted notes, reducing manual rework on complex claims by 31%. Implemented schema-level validation checks in Python to catch missing or conflicting claim attributes, lowering downstream adjudication errors by 18%. Designed a fraud-signal enrichment pipeline to derive temporal patterns, claimant behavior indicators, and location-consistency features from historical claims data, strengthening upstream risk signals. Ran stability checks using stratified sampling and back-testing, keeping fraud recall stable across seasonal claim surges without increasing false positives.

Data Scientist – Transaction Risk & Anomaly Analytics at Space Infolab

March 1, 2020 - July 31, 2023

Analyzed high-volume transaction logs using Python and SQL to flag abnormal spending, velocity spikes, and device-location mismatches, achieving 91% recall on unusual transaction patterns during offline validation. Engineered behavioral features (transaction frequency, geo-drift, merchant diversity) to stabilize risk signals across customer segments and reduce noisy alerts. Handled class imbalance using sampling techniques and probability calibration, cutting false-positive alerts by 27% without lowering overall recall. Evaluated model performance using ROC-AUC, precision–recall curves, and confusion matrices, ensuring consistent behavior across monthly and seasonal cycles. Worked with operations teams to tune decision thresholds based on observed drift patterns, reducing unnecessary escalations by 19% while preserving risk coverage. Participated in code and model reviews to document feature logic, assumptions, and validation results, improving reproducibility and audit readiness.

Applied AI Engineer – Retrieval & Summarization Platform at BlackRock

February 1, 2025 - Present

Built a document retrieval layer in Python using a Pinecone vector index to search filings, portfolio commentary, and research notes, improving Top-3 result relevance by 32% in internal evaluation runs. Improved search accuracy for ambiguous analyst queries by tuning LLM-based query rewriting and BM25 hybrid reranking, increasing Top-1 match rate by 14% for short, vague questions. Added structured coverage checks in Python to enforce required tickers, risk terms, and date windows in summaries, raising summary completeness from 71% to 88% across a 200-sample review set. Set up weekly regression testing on ~500 benchmark questions using RAGAS-style evaluation metrics and a custom financial rubric, blocking 3 prompt/template releases after detecting accuracy regressions. Reduced response latency during market-open spikes by caching high-frequency tickers and briefs in Redis, cutting repeat-request latency by 45%.