A deterministic control plane between an LLM orchestrator and worker agents running in isolated containers. Task state lives in a file-backed vault; workers claim tasks atomically, hold them under lease and heartbeat contracts, and crashed workers are detected by lease expiry and their tasks re-queued. The control plane is stateless, so it restarts without losing work. Modeled on the Kubernetes controller plus etcd pattern.

A drop-in reranking layer that makes RAG retrieval time-aware, even when documents carry no timestamps. Applies explicit half-life decay over retrieved chunks and fuses it with vector similarity, so time-sensitive queries stop surfacing stale context. +191% on fresh-intent queries. Installable with pip install halflife-rag.

The reference application on top of ACORDE: markdown notes, [[wikilinks]], a backlink graph, daily logs and full-text search, with offline-first storage and peer-to-peer encrypted sync across devices. No cloud provider and no accounts.

A microservices-based social platform that replaces traditional engagement metrics with a limited, transferable token supply. Features an MLFQ scheduling-inspired feed algorithm, progressive taxation on token accumulation, and economic incentives for genuine content curation. Built with Next.js frontend, Go/Node.js microservices, PostgreSQL, and gRPC inter-service communication.

Always-Available Conflict-free Offline-first Replicated Distributed Data Synchronization Engine. Built with Go and libp2p, uses CRDTs for automatic conflict resolution in distributed environments. Features XChaCha20-Poly1305 content encryption, SQLite for local persistence, and robust peer discovery. Designed for offline-first apps needing reliable sync without centralized coordination.

From-scratch implementation of the BitTorrent protocol in Go, handling metadata exchange, peer handshaking, piece selection, choking/unchoking strategies, and tracker communication. Implements BEP-3 and BEP-23 compact peer lists, with tests and CI covering bencode, the handshake layout, wire-message framing and info-hash derivation. Peer discovery is tracker-based.

Transforms messy GitHub discussion threads into structured, developer-ready digests. Uses GraphQL to extract issues, PRs, and reviews across a repository, then applies LLM analysis to identify key architectural decisions, rejected alternatives, and constraints buried in comment threads. Features D3.js visualizations for dependency graphs and decision timelines.

Research project on how machine learning can dynamically optimize database index selection based on query patterns and data characteristics. Uses meta-learning to predict optimal index configurations for given workloads, adapting to changing query distributions. Evaluated on standard database benchmarks against traditional cost-based optimizer approaches.

RAG system enabling intelligent document processing and querying. Combines vector embeddings, semantic search, and LLM generation to answer questions from personal document collections. Features document chunking strategies, embedding optimization, context window management, and retrieval pipeline tuning for grounded responses.

A disk-based B+ tree built to scale beyond RAM. mmap I/O, 4KB pages, 256-page batch allocation, and adaptive 2/3 splitting for sequential loads. Insertion with node splitting, deletion with sibling borrowing and merging, point queries and range scans over a leaf-level linked list. Verified against a 6GB index with 95% of the data on disk.

Community platform where IIT Bhilai students share course materials, past papers, lab notes, and study resources. Features course-based organization, upvote/downvote quality ranking, and role-based access control. Built with Next.js 15 App Router, TypeScript, Supabase backend, and Tailwind CSS with shadcn/ui components.

Daily Pokemon guessing game where players identify a random Pokemon through progressive hints. Features animated reveals, streak tracking, shareable results in Wordle-style emoji grids, and a curated Pokemon database. Responsive design optimized for mobile.