Jan 2026 — Present
Madison, WI
Building an Android overlay app that delivers real-time, context-aware security and privacy guidance to adults with intellectual and developmental disabilities (IDD), as part of an IRB-approved HCI study on interdependent security support.
May 2025 — Aug 2025
Atlanta, GA
Engineered an automated RFI/RFP generation pipeline with IBM Docling, Llama 3, and Gemini, cutting proposal turnaround from two weeks to under 24 hours and shipping a full-stack React + FastAPI chat interface to 50+ users.
Jan 2022 — Mar 2023
New Delhi, India
Built a Python and SQL CRM to track 30+ business leads plus an automated contractor-management tool, lifting successful partnerships from 15 to 19 and adding $18K in revenue through better resource allocation.
Sep 2023 — May 2027
Madison, WI
Dean's List (3x) · GPA 3.755 / 4.0. Coursework spanning algorithms, big data systems, AI, and applied quantitative analysis.
Wildlife camera-trap classification pipeline
Microservices system that turns camera-trap photos into structured biodiversity records: a React + Vite frontend plus independent FastAPI services for upload handling, Celery-based async inference (HuggingFace), and observation management over Supabase and Upstash Redis.
Offline AI skin-cancer screening
React + Vite diagnosis workflow (body-area selection, lesion upload, AI summary, follow-up chat) backed by three swappable FastAPI services for classification, orchestration, and a local Ollama LLM, wiring ONNX outputs into patient-friendly explanations.
LoRA distillation + Elasticsearch retrieval
Distilled 895 QA pairs from lecture transcripts with a Qwen2.5-7B teacher, fine-tuned Llama-3.2-1B with LoRA (perplexity 48.85 → 2.49), and shipped a Streamlit chatbot over an Elasticsearch BM25 RAG pipeline benchmarked at Precision@3 0.87.
Production retrieval + generation workflow
End-to-end RAG pipeline for RFI/RFP automation: layout-aware document ingestion, semantic chunking, and vector retrieval feeding an LLM, wired to a React chat UI with token streaming over a FastAPI backend.
High-throughput Kafka ingestion + crash-safe checkpointing
Streams stock ticks from MySQL into Kafka partitions and persists atomic Parquet batches to HDFS with checkpoint-based crash recovery.
Fault-tolerant ingestion with tunable consistency
Ingests NOAA weather data into a 3-node Cassandra cluster (RF=3) over a gRPC API, with tunable read/write consistency and automatic R=2 to R=1 read failover verified through docker-kill fault injection.
Fault-tolerant gRPC address lookup
High-performance address-lookup microservice in Python with gRPC and Protocol Buffers, featuring round-robin load balancing, automatic failover with exponential backoff, and an LRU cache, all orchestrated with Docker Compose.
Spark SQL + ML pipelines + LLM query interface
Analyzes large-scale CodeContests data with Spark RDD/DataFrame/SQL APIs, optimized bucketing/caching, and natural-language query translation.
Airbyte → Snowflake → dbt
Multi-source ELT pipeline ingesting survey, trading-book, and stock/FX time-series data into Snowflake, with 6 dbt staging models and a fact table that validate ticker-date pairs, join point-in-time prices, and compute desk-level profit and loss.
BM25 + kNN vector search on Elasticsearch
Explicit Elasticsearch mappings for 300+ products across 13 categories with a multi-modal query layer (BM25, fuzzy typo tolerance, phrase, boosted multi-field, boolean filters, 384-dim kNN) plus Kibana dashboards for price, ratings, and geospatial density.
MySQL to HDFS pipeline with gRPC failover logic
Builds partition-aware analytics workflows with replication-aware recovery, intelligent caching, and production-style gRPC APIs.
Gradient-boosted demand forecasting
Merged 5 heterogeneous datasets into a unified time-series matrix, engineered calendar, holiday, oil, and transaction features, and compared LightGBM, CatBoost, and XGBoost on a time-based split with Optuna tuning, scored on the Kaggle RMSLE leaderboard.
Investor-startup matching + analytics platform
Investor-startup matching platform over 1,000+ company profiles scraped with Firecrawl and a custom Python parser, with smart filtering on a React + Express + SQLite stack and Recharts dashboards for valuations and funding trends.
Weekly-goal planning & execution dashboard
Next.js App Router productivity dashboard that turns weekly goals into daily action with habit scheduling, task tracking, weighted scoring, daily/weekly/monthly grades, and a Recharts calendar heatmap, backed by Supabase auth and Postgres.
Study-habit tracker + analytics
Study tracker with a dual-mode timer, subject manager, editable session history, and a Recharts dashboard (daily totals, weekly trends, productivity scores, calendar heatmap), with hybrid Firestore + localStorage persistence and per-user security rules.
Multi-threaded processing + GIL benchmark analysis
High-performance concurrent word-count pipeline with configurable thread pools and benchmark suites comparing GIL/no-GIL scaling across CSV, Parquet, and Arrow.
MongoDB aggregation + geospatial queries
30 MongoDB queries spanning multi-stage lookup/unwind aggregation pipelines and a 2dsphere geospatial layer (polygon containment, nearest-neighbor, radial filters), visualized over Santa Barbara city-limit GeoJSON with GeoPandas.
Automation startup website + product surface
Marketing and product site for an automation startup, built with Next.js and TypeScript: responsive landing pages, animated sections, and a reusable component system, deployed on Vercel.
Window functions + moving averages
30 SQL queries across correlated subqueries, multi-table joins, GROUP BY/HAVING, and window functions (RANK, ROW_NUMBER), including a 3-month moving-average revenue analysis over a Dockerized MySQL schema loaded via SQLAlchemy.
Inventory-aware consumer planning app
React + Next.js meal planner that suggests recipes from your on-hand inventory across a 10,000+ recipe API, with dietary filters, saved favorites, grocery lists, and barcode and receipt scanning.
Regression benchmarking + residual analysis
Compares Linear, Ridge, Lasso, Random Forest, and XGBoost regressors after correlation-guided feature selection, log transforms, scaling, and residual plotting.
Unsupervised segmentation + dendrogram analysis
Clusters country-level indicators into interpretable groups using unsupervised learning and dendrogram visualizations to compare development patterns.
Normalized pipeline + train/eval + top-3 output
Feedforward PyTorch classifier for FashionMNIST that reaches 85%+ accuracy in 5 epochs, with a normalized data pipeline, SGD + cross-entropy training, and top-3 Softmax class probabilities at inference.
Posterior classification for English vs Spanish
Implements Bayes theorem over letter-frequency distributions to classify shredded text, with a clear inference pipeline from character counts to language posterior scores.
PCA-based computer vision pipeline
PCA pipeline over 1,000 celebrity images that reduces 3,000-dimensional face vectors into compact eigenfaces, with image reconstruction, eigenface visualization, and noise-perturbation analysis in NumPy and SciPy.
Graph-powered navigation engine
Java routing app over 160+ UW–Madison locations using Dijkstra's algorithm on a hashtable-backed graph, with an HTML interface for shortest paths, travel times, and the 10 nearest destinations from any point.
BST / red-black tree song explorer
Java command-line app that loads song catalogs from CSV and supports attribute filtering (danceability, energy) plus efficient range and recency queries using custom binary search tree and red-black tree implementations.
Skill Atlas
User-facing web interfaces and interaction systems.
Services, APIs, and database systems.
Practical ML and LLM tooling for production workflows.
Data pipelines, distributed processing, and deployment.
I'm Vibhrav Jha, a full-stack and applied-AI engineer studying CS, Data Science, and Economics at UW–Madison. I own features end to end, from interface polish to backend APIs and ML pipelines. I've shipped production software at VDart, built a research prototype at the People & Robots Lab, and delivered 27 projects across web, data, and AI. I'm looking for 2026 SWE and Applied AI roles where I can ship work that reaches real users.
From UX to API architecture
I ship complete features across React frontends and FastAPI or Node backends, with an emphasis on user clarity, clean abstractions, and maintainable delivery.
Production systems, not demos
I build AI workflows backed by numbers: 96% retrieval accuracy on a production RAG pipeline, and a 20x drop in model perplexity from LoRA fine-tuning.
Reliability and communication
I ship fast without cutting corners: deterministic safety floors, fault-injection testing, and CI/CD on every push, with decisions documented as I go.
Based in Madison, WI, originally from New Delhi, India.
High-quality product engineering with real user impact. Switch the globe to Places I've been and tap a flag to spin there.
Outside work
New Delhi to Madison. 12,000 km from where I grew up to where I build now.