Things I've built

A production-grade, self-healing SRE control plane built as a 6-agent swarm over NATS JetStream. Monitors 10+ polyglot microservices, performs LLM-assisted root cause analysis, and executes safe remediations with human-in-the-loop approval.

A lightweight, transactional, SQL-compliant database engine built from scratch in Go. LSM-tree storage with MemTable + SSTables, WAL durability, ACID transactions via MVCC + snapshot isolation, background compaction, hash indexing, and a custom SQL parser. Comes with an interactive REPL and a TCP server mode.

A live, interactive visualizer for the Raft consensus algorithm. Simulates a 5-node cluster with a virtual networking layer; lets you deterministically inject faults (partition networks, kill leaders, restart nodes) and watch the cluster re-elect, replicate logs, and heal in real time. Go backend, React frontend.

A browser-based, interactive simulator for the Hierarchical Navigable Small World algorithm — the graph-based ANN structure behind Qdrant, Weaviate, pgvector, and FAISS. Click-to-insert nodes, click-to-search, watch the layered descent, isolate any layer, and inspect a structured log of every operation.

A real-time collaborative document editor where multiple users edit simultaneously without conflicts. Custom-built RGA (Replicated Growable Array) Sequence CRDT — no external CRDT libraries, no Operational Transformation. Handles out-of-order ops, tombstone deletes, ghost-delete caching, and live remote-cursor presence. WebSocket transport, Monaco editor, SQLite persistence.

An educational 32-bit x86 OS written in C and NASM assembly. Multiboot-compliant boot via GRUB, GDT + TSS + IDT setup, ring 0/3 privilege separation, two-level paging, kernel heap, PIC remapping, VGA text-mode, PIT, PS/2 keyboard, in-memory ramfs, round-robin scheduler with context switching, 10 syscalls via int 0x80, and an interactive userspace shell — all with zero external dependencies (no libc, no standard library).

An MCP (Model Context Protocol) server that lets LLM orchestrators automate GitHub pull request workflows safely. Isolated sandbox execution, command allowlisting, fork-only workflow, multi-layer approval gates, secret redaction, and repo allowlists.

A GraphRAG system that combines vector search (Qdrant) with knowledge-graph reasoning (Neo4j) to deliver explainable portfolio insights. Constructs a portfolio KG, performs hybrid retrieval, computes sector exposure, reasons about event impact, and returns structured answers with citations.

A production-ready customer support triage system that classifies tickets, retrieves relevant policies, drafts responses, and enforces compliance — all with human-in-the-loop escalation. Four agents work in sequence with LangSmith tracing for observability.

A multi-agent arXiv pipeline that intelligently retrieves, processes, and synthesises research papers. Implements adaptive retrieval, self-correction loops, and multi-agent coordination via LangGraph. Routes queries, scores relevance, generates citations, and self-evaluates outputs.

A multimodal deepfake analysis system integrating CNN/Vision Transformer detection, explainable-AI techniques (SHAP, LIME), and image-to-text captioning with CLIP and attention mechanisms. An LLM contextual explanation module using Llama 3.1 + Ollama produces forensic narratives. Deployed as REST APIs for integration into forensic workflows.

A mobile app that connects entrepreneurs with potential co-founders. A weighted scoring algorithm evaluates compatibility across skills, interests, goals, startup stage, location, availability, and collaboration style, then surfaces match recommendations with explanations.

A web-based stock analysis platform that helps investors value stocks and discover growth opportunities across thousands of U.S. equities. Provides search, screening, watchlists, price alerts, technical analysis, and valuation tools. Built as a microservices architecture on AWS.

ML pipeline using DeiT-small to classify MRI scans as Alzheimer's Disease or Normal Control. 81.42% validation accuracy. Implemented patch extraction, linear embeddings, positional encoding, and multi-head self-attention from scratch in TensorFlow/Keras, trained on ADNI.