AI Engineer · Data Scientist · New York, NY
Hey, I'm Sarvar — an AI engineer and data scientist in NYC who loves building things that make people's work easier (open to relocate).
Most recently I built Siftdog, an open-source, self-hosted web search API for AI agents. It's a drop-in Tavily replacement, runs as an MCP server, and in my benchmark matched Tavily on factual answers at $0 per search. Code's on GitHub.
I also founded AcadyLearn, an AI app that turns your documents into quizzes and flashcards. I built it end-to-end, from the LLM workflows to the AWS infrastructure.
A lot of my work has been in healthcare and clinical research. At AbbVie I built agentic systems for statistical workflow automation: LLM pipelines, document Q&A, and a full-stack analysis platform biostatisticians use day-to-day.
Before that, at EDA Clinical I built clinical data quality models and risk indicators across Phase II–III studies, plus models predicting enrollment and patient dropout. Earlier, at Penn State, I built computer vision pipelines that spotted overcrowding in transit video.
I've done ML research at the Feng Lab too, pitting CNNs against multimodal models like CLIP and Qwen-VL on image classification.
At a hackathon I built HealthLab Agent, an agentic workspace for exploring public health data from the CDC. More projects here and on GitHub.
AcadyLearn
Built and shipped a production AI application that transforms uploaded documents into quizzes and flashcards. Designed end-to-end LLM workflows with asynchronous job orchestration, fault-tolerant background processing, and cloud infrastructure on AWS.
AbbVie
Developed a GenAI governance review workflow for LLM-based clinical and scientific document summarization, cutting manual intake review time by 40%. Built and validated NLP pipelines to extract entities, adverse-event signals, and study endpoints from unstructured clinical documents, and deployed ML models in Python and SQL to predict patient enrollment risk, site activation delays, and protocol deviations. Ran quantitative LLM evaluations across accuracy, groundedness, and safety metrics, improving acceptable response rates by 20–35%.
EDA Clinical
Built clinical data quality models across Phase II–III studies to detect anomalous lab values, visit-window deviations, and site-level risks, reducing manual query review by 30%. Developed ML-based risk indicators for centralized monitoring by combining EDC, CTMS, lab, and eCOA data, and shipped automated review dashboards and edit-check routines that improved early detection of data discrepancies by 25–40%. Designed reusable SDTM-like data pipelines and predictive models for enrollment forecasting and patient dropout.
Penn State
Developed an end-to-end OpenCV–TensorFlow computer vision pipeline trained on 1.5 years of transit video data to detect overcrowding patterns and support a 10% reduction in congestion. Curated and augmented large-scale video datasets across varying lighting, camera angles, and crowd densities, and trained detection models (ResNet, Inception, EfficientNet, EfficientDet, YOLO), reducing false detections from 15% to 0.1%.
Feng Lab
Evaluated lightweight CNNs and ResNet architectures against multimodal models including CLIP and Qwen-VL on TinyImageNet. Fine-tuned CLIP ViT-B/32 for a 10% accuracy gain over zero-shot baselines.
New York University
Graduate study in biostatistics with a focus on machine learning and statistical modeling for health data.
Pennsylvania State University
Undergraduate study in data science covering algorithms, applied statistics, and computational methods.
A self-hosted web search API for AI agents and LLM apps that works as a drop-in replacement for Tavily: the same /search, /extract, /crawl, and /map endpoints, so the official Tavily SDK and LangChain integration work by changing only the base URL. Uses SearXNG metasearch, trafilatura extraction, and BM25 ranking, with optional Claude-written answers, and ships as an MCP server for Claude Code and Cursor. It needs no search API key. In a reproducible 45-query benchmark it matched Tavily on factual answers (25/25) and found official documentation more often (9/10 vs. 3/10 in advanced mode), at $0 per search.
Can a 500M-parameter vision-language model be meaningfully fine-tuned for K–8 science MCQ under strict hardware constraints (5M trainable params, free-tier T4 GPU)? Applied QLoRA with four staged ablations covering rank, alpha, target modules, and scoring method. Reached 0.875 public leaderboard accuracy, up from a 0.819 baseline — with a key finding that LoRA alpha interacts with module scope: raising alpha hurts attention-only adapters but helps when MLP projections are also included.
An agentic workspace for public health data exploration built at a hackathon. Users upload CSVs or pull datasets directly from the CDC/Socrata catalog, then run LLM-driven loops for profiling, cleaning, and analysis — generating charts, statistics, and Markdown reports. Integrates PubMed to surface relevant literature alongside the data findings.
Can ML models reliably flag high-risk pregnancies from CTG signals — reducing reliance on operator interpretation? Compared logistic regression, Lasso, and random forest on 2,126 CTG records (UCI) classifying fetal health as Normal, Suspect, or Pathological. Random forest achieved 94.6% accuracy and 96.9% balanced accuracy on the critical Pathological class, with abnormal short-term heart rate variability as the top predictive feature.
Does antihypertensive medication use change how depression affects blood pressure control? Modeled the interaction using multivariable logistic regression and random forest on 39,467 NHANES participants (2005–2020) with survey-weighted analyses. Among medicated patients, each unit increase in depression score raised the odds of uncontrolled hypertension (aOR: +0.04, 95% CI: 0.02–0.06); logistic regression slightly outperformed random forest on AUC (0.79 vs. 0.78).
Do states with higher inactivity rates see faster obesity growth over time? Applied linear mixed-effects models to 709 state-year observations from CDC BRFSS and Census ACS data (2011–2024). Obesity rose ~0.58 pp/year on average; 67% of total variance was attributable to stable between-state differences (ICC = 0.67). Physical inactivity and poverty both independently predicted higher obesity prevalence after controlling for time trends and clustering.
Does the open-source 50-item Big Five test hold up at scale? Evaluated reliability and validity using a 100,000-observation random sample from 1M+ international respondents (2016–2018). Cronbach's alpha ranged from 0.79–0.89 across all five traits, exploratory factor analysis with parallel analysis reproduced the expected five-factor structure, and low inter-trait correlations confirmed divergent validity.
How did COVID-19 mortality risk vary across U.S. states over time? Calculated monthly CFR for all 50 states using NYT surveillance data, then visualized trends with faceted line plots and choropleth maps. CFR declined nationally across 2020–2023, with northeastern states peaking highest early in the pandemic.
Do the genres you listen to predict how depressed you feel? Built a multiple linear regression model on 728 respondents from the MxMH Survey (Kaggle) using genre frequencies and demographics as predictors. The model explained 18.2% of variance in depression scores; age was protective, Classical listening was positively associated with depression, and Country showed a significant negative association.