Deepesh Kumar

AI document pipelines

• Built a daily legal-monitoring pipeline that scrapes SAFLII for new Government Gazettes, flags tracked corporate Acts, rewrites them into plain-English summaries via the OpenAI Batch API (50% cheaper than standard calls) and files Word reports to SharePoint, all scheduled on GitHub Actions.
• Built LegalProcessor, an 8-stage PDF → plain English → DOCX pipeline publishing to Azure Blob, SharePoint and WordPress, with a React + TypeScript dashboard that tracks every stage live.

Outreach & scheduling

• Built a multi-tenant lead-engagement platform (FastAPI, Postgres, Microsoft Graph) automating Outlook outreach, reply tracking, Teams bookings and Playwright-driven LinkedIn outreach, with human-approved AI replies and an MCP server for internal assistants.
• Shipped a WhatsApp meeting-reminder service (Node.js, Prisma, Twilio) that syncs Microsoft 365 calendars through Graph webhooks and messages every attendee 10 minutes before a call, handling reschedules, cancellations and duplicate-safe sends.
• Built a real-time office café ordering app (Next.js, Prisma, SSE); baristas run the queue from WhatsApp buttons.

AI monitoring & insights

• Built multi-LLM brand monitoring across 6 providers with weekly mention and ranking reports (Flask, MongoDB, Next.js dashboard), contributing to a 21% rise in brand visibility across major AI models in 4 months.
• Built YouTube audience intelligence that turns comments across the channel’s 294 videos into ranked video ideas, content gaps and unanswered questions with gpt-4o-mini, at about $0.10–0.20 per run.

Engineering tooling

• Built an incident copilot (FastAPI, ARQ/Redis, Postgres) that ingests production alerts, pulls context from GitHub, Vercel, Azure and Forge, writes AI triage, routes Outlook escalations to service owners and files reports to SharePoint.
• Building a PR code-review bot on an internally hosted LLM so code never leaves company infrastructure; now in a silent-mode pilot.

Product build

• Rebuilt the core web app’s architecture and user flows, shipping a stable production release in 1.5 months that lifted conversion and cut critical errors.
• Automated bank-statement parsing and added progress tracking and engagement analytics, cutting manual intervention by 45%; set up CI/CD and Husky.

Real-time signal tools

• Built a biosignal visualizer (Next.js, Django) for EMG, ECG, EEG and EOG, streaming 500+ points/sec over WebSockets with automatic serial-port detection.
• Designed Django APIs for multi-channel FFT analysis, plotting and data export across 8+ channels.