Sanskar Shukla

A production LLM application: six providers behind one abstraction, schema-validated outputs, retry/timeout/circuit-breaker fault tolerance and persistent cross-game agent memory.

A real-time multiplayer poker platform where humans play Texas Hold'em against Claude, GPT, Gemini and Grok every decision is a live LLM call with pot odds math, opponent reads, trash talk, and persistent memory of how you play.

Key Features:

• Server authoritative game engine in pure TypeScript immutable state transitions for betting rounds, showdown evaluation across all 21 five-card combinations of 7 cards, and full multi-way side-pot/split-pot distribution clients send intents only, so cheating is architecturally impossible
• Per-player state masking over WebSockets a single Socket.io room where each connected client receives an individually masked game state, guaranteeing hole cards never leave the server two-tier state (in-memory hot path + Redis persistence) lets live games survive server restarts and redeploys mid-hand
• BYOK credential vault users' AI provider keys encrypted at rest with AES-256-GCM using the owner's user ID as authenticated data, making ciphertexts cryptographically non-transferable between accounts keys are never returned by any API or written to logs
• Six providers, three wire protocols, one abstraction Anthropic, Google, and OpenAI compatible adapters with per-seat OpenRouter routing; adding a new OpenAI compatible provider is a single catalog entry with zero adapter code
• Fault tolerant LLM orchestration every model call runs inside a timeout race with schema validated parsing, retry with backoff, and a per model circuit breaker, so a failing provider degrades to safe auto-actions without ever freezing the table
• Cross-game AI memory models persist opponent profiles, behavioral stats (VPIP, aggression), and free-form reads to Postgres between sessions play Claude twice and it opens with a scouting report on you

An AI workflow engine: LLM nodes (Anthropic, OpenAI, Gemini) run inside a dependency-ordered execution graph with encrypted credentials and real-time run tracing.

A visual workflow automation platform where you drag nodes onto a canvas, wire them together, and execute multi-step pipelines that chain HTTP requests, AI model calls, and third-party integrations with zero glue code.

Key Features:

• Drag-and-drop workflow editor built on ReactFlow with real-time canvas state synced to PostgreSQL through a delete-all-then-recreate transaction pattern that preserves encrypted secrets across saves, so credentials never leave the server unencrypted
• Six AI provider nodes across Anthropic, OpenAI, and Gemini with a unified credential vault where API keys are AES-encrypted at rest using per-node secret persistence, and OpenRouter routing lets users swap models without touching workflow logic
• Topological execution engine that resolves node dependencies from the connection graph, runs each node in dependency order with a 60-second per-node timeout and abort controller, and writes partial output on failure so you can debug exactly where a pipeline broke
• Template variable system with {{variableName}} interpolation across nodes, where each node's output is namespaced by its variable name and flows downstream as context, making multi-step API chains composable without hardcoded values
• Trigger system supporting manual execution, inbound Google Form webhooks with HMAC-verified secrets, and Stripe webhook events with configurable event type filtering and signing secret validation
• Discord and Slack integration nodes that post formatted messages to channels with full context from upstream nodes, plus HTTP request nodes with redirect following, DNS pinning against SSRF, and response size limiting
• Credential management with encrypted storage, per-user isolation, and type-safe provider validation so a Gemini key can't accidentally land on an Anthropic node, with the raw key never returned by any API endpoint or written to logs
• Full execution history with per-run status tracking, timestamped output snapshots, and real-time node status updates streamed through Inngest's realtime channels so the UI reflects execution progress as it happens

A distributed real-time crypto paper trading platform that streams live market data from Binance with sub-50ms latency, supporting margin trading with up to 100x leverage risk free.

Key Features:

• Real-time WebSocket price feeds processing 50K+ market data points per minute from Binance live streams with sub-50ms update latency
• Event-driven architecture using Redis Pub/Sub and Kafka for horizontal scalability and fault-tolerant event distribution across services
• CFD trading engine with margin calculations, leverage support up to 100x, partial profit/loss mechanisms, and automatic position liquidation logic
• Persistent order and position storage with PostgreSQL and TimescaleDB, with in-memory order books and periodic snapshots for crash recovery
• Full trading dashboard with live charts, real-time P&L tracking, and responsive UI built with TailwindCSS for active trading
• Containerised deployment with Docker and Caddy for reverse proxying and automatic HTTPS

A decentralized equipment tracking system that uses the Solana blockchain plus the MERN stack to maintain transparent ownership history for physical assets.

Key Features:

• Role-based access (Manufacturer/User) with JWT-secured auth flows and separate dashboards
• Manufacturers can register equipment with images and documents, then transfer ownership via wallet address
• Every transfer is recorded on-chain to create an immutable ownership timeline while metadata is kept off-chain in MongoDB
• Dark-mode dashboard for managing pending, accepted, and rejected transfers with clear status filters
• Scalable backend with file uploads, validation, and modular APIs ready for NFT-based and QR-code extensions