What I've Built

AI-powered medical fact-checker that breaks any AI-generated answer into atomic claims, verifies each against real clinical guidelines (NICE, AHA), and returns a trust score, hallucination rate, and corrected source-backed rewrite.

• ●3-node LangGraph pipeline (Claim Extractor, Retriever, Scorer) over ~844 guideline chunks from NICE NG28, NG136, CG181, and AHA 2025 CPR; embeddings via all-MiniLM-L6-v2 with cosine similarity search in ChromaDB.
• ●Live web augmentation via Tavily API fires when corpus similarity falls below 0.35, pulling from NIH, WHO, PubMed, Mayo Clinic, and NICE; results shown with a live web badge alongside guideline citations.
• ●Per-claim SUPPORTED / UNSUPPORTED / CONTRADICTED verdicts with rationale and page-level citations; aggregated into a trust score (0-100), hallucination rate %, and source coverage %, validated end-to-end with RAGAS evaluation.
• ●Corrected answer rewrite grounded in retrieved passages with clickable citations; multi-turn follow-up chat persisted per evaluation and shareable by URL with no account required.
• ●Cut backend memory by 70% by switching from PyTorch-based SentenceTransformers to ONNX-quantized FastEmbed embeddings, keeping retrieval quality intact at a fraction of the resource cost.
• ●Full-stack: FastAPI + SQLAlchemy backend deployed across Render and Neon serverless PostgreSQL, React 18 + TypeScript + Tailwind frontend on Vercel, with health checks, automatic restarts, and push-to-deploy continuous deployment; evaluation history sidebar with rename and delete.

Built an AI-powered DevOps agent that monitors production deployments in real time, scans for failures and security issues, and autonomously generates and opens GitHub pull requests with fixes. Integrated with Slack for operator-facing slash commands.

• ●Live deployment monitoring via Vercel webhooks in real time; automatically detects build failures, runtime errors, and security vulnerabilities and triggers the fix pipeline without human intervention.
• ●Autonomous PR generation using the GitHub API: the agent diagnoses the root cause, writes a patch, and opens a pull request with a structured description ready for review.
• ●Slack integration with slash commands for rollbacks, incident queries, and system status checks, giving teams a natural language interface to the agent.
• ●Powered by Gemini 3 Flash via the Vercel AI SDK for diagnosis and code generation; Supabase for persisting incident history and deployment state.

Built a multi-task learning model using BioBERT v1.1 to extract biomedical entities from 5 benchmark corpora. Compared single-task vs multi-task approaches to find what works best for low-resource biomedical datasets.

• ●Achieved F1=0.98 on BC2GM gene recognition and F1=0.99 on NCBI Disease extraction.
• ●Benchmarked across 5 corpora (BC2GM, BC4CHEMD, BC5CDR, JNLPBA, NCBI Disease) using IOBES tagging.
• ●Multi-task model showed stronger cross-dataset generalization on smaller datasets vs single-task.
• ●Engineered preprocessing with token alignment and dynamic batching for efficient fine-tuning.

Built an image captioning system combining InceptionV3 for visual features with a dual-mode Transformer decoder. Trained on 120K+ images and deployed as a real-time Streamlit web app.

• ●Achieved BLEU score of 0.80 on MS-COCO evaluation set.
• ●Trained on 120K+ labeled images across MS-COCO, Flickr8k, and Flickr30k datasets.
• ●Dual-mode decoding: greedy search for speed, beam search for caption quality.
• ●Deployed as a Streamlit web app for real-time caption prediction from uploaded images.

Used multiple causal estimators on 10K+ education records to estimate the real treatment effect of growth mindset programs, not just correlation.

• ●Analyzed 10K+ observational education records modeled after the National Study of Learning Mindsets.
• ●Applied 5 causal estimators: Causal Forests, IPW, S-Learner, T-Learner, and X-Learner.
• ●Found consistent positive causal effect across all methods, confirming estimator-agnostic robustness.
• ●Built propensity score models to control for confounders in observational data.

Combined regression modeling, clustering, and location-based analysis to identify pricing strategies that maximize host revenue.

• ●Identified optimizations projecting a 25% revenue increase for underperforming listings.
• ●Applied regression, clustering, and geospatial analysis across listing features.
• ●Built interactive Tableau dashboards for exploring pricing patterns by neighborhood.
• ●Processed multi-modal data: pricing, reviews, amenities, and geographic coordinates.