ShubhamTiwari

A 12-agent platform that finds, tailors and applies to the best-fit roles.

• Architected a distributed job-automation platform with 12 AI agents — job finder, LinkedIn/Indeed fetcher, résumé tailor, salary negotiator and auto-apply — running on 10–15 async Celery workers backed by Redis.
• Auto-discovers and applies to the highest-fit roles each run (~8–10 scoring 85+/100), rewrites the résumé per job description before submission and tracks every application — with Stripe billing and local Ollama inference.

Event-driven market platform with sub-250ms streaming and ML price signals.

• Engineered an event-driven trading platform streaming live market data over WebSockets into OHLC candles at sub-250ms latency; trained ML models on 4 years of 5-minute data for 150+ companies, fusing a time-aligned news API as a sentiment signal to predict price direction.
• Built an automated risk engine — dynamic position sizing, confidence scoring, stop-loss and drawdown protection — validated through historical backtesting and exposed via modular REST and WebSocket APIs for tracking, predictions and retraining.

Rebuilding Apache Kafka's core from first principles, milestone by milestone.

• Implemented a segmented, append-only commit log with offset indexes, topics and partitions, and a custom binary wire protocol served by a TCP broker — with producer and consumer client libraries.
• Added consumer groups with a broker-side coordinator — partition assignment, generation-based rebalancing, heartbeats and committed offsets — plus a live stream-visualisation UI. Next up: idempotence, compaction, replication and Raft.

Native D3D12 engine & editor that AI agents can drive through 60 MCP tools.

• Windows engine with a docked ImGui editor: D3D12 rendering with GGX lighting, PCF and optional DXR 1.1 ray-traced shadows, GPU instancing and LODs, fixed-step Jolt physics, an EnTT scene model with undo/redo, and standalone game export.
• An MCP adapter exposes 60 strict tools so AI clients can author, test and export games, alongside an async AI assistant (OpenAI-compatible with Ollama fallback), executable behaviour graphs and integrity-checked, bounded asset caches.

Stripe/Razorpay-style gateway with idempotent payments and crash-safe queues.

• Spring Boot API accepts payments in milliseconds (202 Accepted) and settles them asynchronously through Redis Streams consumer groups against 30 simulated banks with realistic latency and decline rates.
• Idempotency keys guarantee one attempt and one receipt per payment under concurrent retries; worker crashes, bank timeouts and Redis outages are absorbed via message reclaim, requeue, dead-letter queues and database fallback — verified by a multi-threaded load generator.

AI that QA-tests web apps like a senior engineer, then files the tickets.

• Built a 7-package monorepo where AI agents autonomously crawl web apps, detect JS errors, broken flows and accessibility issues (axe-core), then auto-generate bug reports and Playwright regression tests.
• Multi-tenant SaaS with Jira/GitHub ticket filing, Stripe billing tiers, and swappable LLM backends (OpenAI, Anthropic, Ollama).

Multi-agent autonomous trading system for the Indian markets (NSE).

• Designed a 4-agent pipeline — Scout (screening), Market Analyst (live candles), Risk Manager (9-point validation) and Trade Executor — coordinated over a Redis event bus.
• Powered decisions with local LLMs (Ollama qwen3) + GPT-4o-mini; a live Next.js dashboard streams trades and candle updates over WebSockets.

Speak a technical concept — watch an accurate, animated diagram build itself.

• Agentic pipeline: Whisper transcription → provider-agnostic LLM gateway (OpenAI / Ollama) → a validated Scene DSL → deterministic GSAP + SVG animation, giving reproducible, editable output instead of black-box AI video.
• A Scene Director splits narration into beats and paces each by its spoken length, with validate-and-repair passes and real-time WebSocket streaming of generation progress.