The problem. Students ask the same questions over and over, and generic AI chatbots either hallucinate or pull from the open internet. Instructors can't trust them to answer only from what was actually taught in the course.

My solution. A RAG-powered AI teaching assistant that answers exclusively from instructor-uploaded materials, with source citations and a strict similarity threshold so it never makes things up.

• Delivers source-cited, zero-hallucination responses via a strict 0.85 cosine similarity threshold across 8 document formats.
• Designed a semantic response cache using pgvector (0.12 distance threshold) that cuts repeated query latency from ~2.5s to ~25ms, a 100x improvement, with no extra LLM calls.
• Built a multimodal ingestion pipeline that extracts and captions images from PDF (OCR), PPTX, and DOCX using Gemini Flash vision at upload time, plus a cross-session memory layer injecting the top 3 prior Q&As (within 0.55 cosine distance) into each query.
• Architected a fully async FastAPI / asyncpg / SQLAlchemy backend with executor-offloaded embeddings, hardened by 12-pattern prompt injection guardrails and 5 req/min rate limiting, and a privacy-first analytics dashboard tracking 8 question themes with aggregated-only metrics. Private repo, working end-to-end prototype.

The problem. Job hunting at volume means copy-pasting dozens of postings into spreadsheets, extracting details by hand, and losing track of where each application stands. Spreadsheets stop scaling fast.

My solution. An AI job tracker built across 10 sprints: paste up to 5 job descriptions at once and Gemini AI extracts company, role, salary, skills, and H1B sponsorship in a single batch call.

• Scores your resume against each job on a 0 to 100 scale using a calibrated senior recruiter persona, and generates 10 tailored interview questions.
• Dashboard surfaces real conversion rates, weekly trends, and stale applications across three view modes (cards, table, Kanban). Built on Next.js 14 (Vercel), FastAPI (Render), and PostgreSQL (Supabase). Private repository.

The problem. A stock's daily price is just noise until you can place it in context: which moments actually marked a structural shift, what was really driving them, and where the market was simply drifting.

My solution. Analyzed 1M+ records across 25+ years (1999–2025) of a leading chipmaker's stock to pinpoint the AI-driven inflection points, then told the story in an interactive Tableau dashboard.

• Identified 2016 and 2023 as breakpoints where price grew ~30× from pre-AI levels (under $5 to $149+), with volume surging 56% at the 2017 deep-learning inflection.
• Built a 2-page interactive Tableau dashboard (year toggler, price-range overlay, closing-price trends, volume breakdowns) enabling comparison across three distinct market eras.

The problem. Airlines plan capacity months ahead, but raw historical traffic is too big and too noisy to forecast from directly. It has to be cleaned, aggregated, and modeled before it can guide a seat-planning decision.

My solution. Processed 3.3M airline records with Apache Spark and trained an LSTM neural network that forecasts Delta's 2025 passenger volume at about 88% accuracy.

• Paired the model with four Tableau dashboards covering quarterly trends, passenger patterns, and cargo analysis, plus a full report comparing predictions vs. actuals.