Kumar Deepansh
AI agents & full-stack systems, shipped to production
Overview
Lead Engineer, AI Voice Platform @Assistable
Founder @Lumbox
Software Development Engineer & Team Lead @XAMTAC CONSULTING LLC
Chicago, IL / Remote
Social Links
Work with me
Available for contract work, 20 to 25 hours a week, overlapping US mornings and the UK afternoon.
- AI features and agents that hold up in production: tool calling, RAG, structured outputs, guardrails and tracing, added to your web app. Fixed-price projects from $500.
- MCP servers: a token-lean server for your API, or an audit that cuts what your existing one costs an agent. Audits from $350.
- Full-stack builds and rescues: Next.js, Node/TypeScript, Python and Postgres. I take over existing codebases and make them reliable.
- Backend scaling: queues, background jobs, slow queries and load testing, with a written report of what I found and fixed.
About
I build AI features and the production systems they run on.
For the last year I have been the sole engineer on an AI voice and messaging platform: 300,000+ sub-accounts, 13.3M+ calls, 100M+ messages. I rebuilt it from Bubble.io onto Next.js, Node/TypeScript and Postgres, which cut infrastructure spend 60%. I built the queue runtime behind every AI workflow (40+ queues, load-verified at 1,000 req/s) and added Langfuse tracing on the LLM calls, so a failed run can be explained, not guessed at.
Before that, four years at an AI and software agency, shipping client work from scope to deploy: dashboards, client portals, multi-tenant SaaS, and an AI marketing ERP with 1K+ daily users. I became Team Lead over 14 engineers there.
On my own time I build agent infrastructure: Lumbox gives AI agents a mailbox, a browser session and a credential vault behind a 77-tool MCP server, and AgentCursor lets an agent drive any Mac app by reading windows as text instead of screenshots.
A Smart India Hackathon winner, with IPR for EdTech work used by 10,000+ students.
Stack
Blog
The Right Way to Move Array Elements in JavaScript and React
- Published on
AI Agent Frameworks: The Best AI Agent Tools in Python, Node.js, and Go
- Published on
Unlocking the Future: The Power of Vertical Large Language Models
- Published on
Tame Your Terminal: 3 Foolproof Ways to Stop Conda Automatic Activation
- Published on
Experience
Assistable
Current Employer- Sole owner of a multi-tenant AI voice and messaging platform: 300,000+ sub-accounts, 13.3M+ calls and 100M+ messages lifetime, 10M+ events/month. Product code, data model, infrastructure, deploys and on-call
- Migrated 57,000+ AI voice agents onto a Telnyx telephony backbone at under 200ms RTT and 14% lower cost; built number and SIP provisioning, WebRTC token issuance, an idempotent call-event inbox and a containerized post-call worker runtime
- Replaced a Bubble.io no-code platform with a Next.js + Node/TypeScript rebuild on Postgres, cutting infrastructure spend 60%
- Built the automation runtime behind every AI workflow: 40+ concurrent queues driving a worker fleet that autoscales on queue depth, with dynamic batching and backpressure bounded by Postgres pool headroom. Load-verified at 1,000 req/s sustained
- Instrumented the AI stack with Langfuse LLM tracing, a ClickHouse event store and 10+ Grafana dashboards; led response and postmortem on a 9-hour ingest outage, root-caused to a promise leak in the flush path
- TypeScript
- Next.js
- Node.js
- PostgreSQL
- Voice AI
- Telnyx
- WebRTC
- Queues
- Langfuse
- ClickHouse
- Grafana
Lumbox
Current Employer- Building Lumbox, email infrastructure for AI agents: a programmable inbox agents read, send, and act on via a single API and an MCP server exposing 70+ tools
- Designed and shipped a full email stack (sending, receiving, threading, search, contacts) plus an OAuth 2.1 authorization server for secure agent access
- Built a credential vault (AES-GCM, per-org keys) and TOTP 2FA so agents can safely log in to third-party services
- Operate self-hosted mail infrastructure (Haraka SMTP, Dovecot IMAP, DKIM/SPF/DMARC, BIMI) with deliverability tuning
- AI Agents
- MCP
- Email Infrastructure
- OAuth 2.1
- TypeScript
- Cloudflare Workers
- PostgreSQL
- SMTP/IMAP
- System Architecture
- Founder
XAMTAC CONSULTING LLC
Current Employer- Spearheaded creation of AI-powered marketing ERP system (app.xamtac.com), driving 25% increase in client conversion rates and supporting 1K+ daily active users
- Architected key features of Xamtac SaaS platform, driving 35% increase in adoption and 30% boost in client retention: • Engineered asset library system and multi-channel campaign tool, reducing time-to-market by 40% • Implemented PPC reporting and CRM functionalities, increasing lead conversion by 30% • Designed client portal and email marketing suite with no-code builder, elevating email engagement by 50% • Integrated social media platforms with cross-platform scheduling, cutting workflow time by 70%
- Orchestrated CI/CD pipelines and automated testing, reducing deployment time by 40% and mitigating production bugs by 60%
- Promoted to Team Lead, mentoring and guiding a 14-member development team, elevating overall code quality by 35%
- Directed creation and execution of key client deliverables for high-value clients like Poshmom E-Commerce and TalkTales
- React
- Next.js
- Node.js
- TypeScript
- Python
- PostgreSQL
- MongoDB
- AWS
- Docker
- CI/CD
- Team Leadership
- Project Management
- System Architecture
Rebolt (via XAMTAC CONSULTING LLC)
- Leading Xamtac team as Team Lead for Rebolt's full-stack development initiatives
- Architecting and implementing comprehensive test suites to ensure code quality and reliability
- Developing automated ads system for Rebolt's clients, optimizing campaign performance and ROI
- Managing end-to-end full-stack development using modern technologies
- Collaborating with cross-functional teams to deliver scalable solutions
- Full Stack Development
- Test Suite Development
- Automated Ads Systems
- Team Leadership
- React
- Node.js
- TypeScript
- Test Automation
- Ad Tech
- System Architecture
The Volte
- Optimized e-commerce platform, reducing JavaScript bundle size by 40% and improving core web vitals by 130%
- Engineered responsive UI components with Next.js, boosting mobile user engagement by 45%
- Implemented advanced SEO strategies, driving 70% increase in organic traffic within two months
- Next.js
- React
- JavaScript
- TypeScript
- Performance Optimization
- SEO
- Responsive Design
RAAHEE
- Created a mental health platform for booking therapist sessions, processing 500+ daily transactions with optimized RESTful APIs
- Designed a real-time availability calendar, reducing booking conflicts by 60% and streamlining the appointment process
- React
- Node.js
- Express.js
- MongoDB
- RESTful APIs
- Real-time Systems
Projects(18)
Founder. Email infrastructure built for AI agents: a programmable inbox that agents can read, send, and act on through a single API and a Model Context Protocol (MCP) server exposing 70+ tools.
- Full email stack: sending, receiving, threading, search, and contacts
- MCP server with 70+ tools so any agent can run an inbox out of the box
- OAuth 2.1 authorization server for secure agent access
- Credential vault (AES-GCM, per-org keys) and TOTP 2FA for agent logins
- Self-hosted mail stack (Haraka SMTP, Dovecot IMAP, DKIM/SPF/DMARC, BIMI)
- AI Agents
- Email Infrastructure
- MCP
- OAuth 2.1
- TypeScript
- Cloudflare Workers
- PostgreSQL
- SMTP/IMAP
- API Design
Open source (MIT) MCP server and TypeScript SDK that lets an AI agent drive any Mac app or browser tab with a visible, human-like cursor.
- Reads windows through the macOS accessibility tree as compact text: 60 to 130 tokens a window, against 600 to 1,200 for a screenshot
- Clicking a button by label costs about 9 tokens, which makes computer use practical for local models
- A 27B model running entirely on a laptop completed a read, click, type and verify loop through it with no network calls
- One setup command connects Claude Code, Cursor, VS Code, Codex, Windsurf, Claude Desktop and Gemini CLI
- AI Agents
- MCP
- Computer Use
- TypeScript
- Swift
- macOS Accessibility
- Open Source
Local-only health and activity tracker for Android phone and Wear OS (Galaxy Watch 7). The watch streams raw sensor data to the phone, which runs a research-backed calorie and sleep engine entirely on-device - no servers, no accounts.
- Wear OS app captures accelerometer + heart rate with adaptive sampling
- Phone calorie engine using Mifflin-St Jeor BMR, Hildebrand accelerometer, and Keytel HR equations
- On-wrist + real-movement detection so only genuine active calories are counted
- Sleep detection from wrist actigraphy (Webster rescoring)
- Custom Bluetooth LE bridge between watch and phone (iOS phone path planned)
- Kotlin
- Jetpack Compose
- Wear OS
- Android
- Bluetooth LE
- Room
- Gradle
- Sensors
- On-device Health
Free, private, in-browser video compressor. Compresses and converts video entirely client-side with WebAssembly, so files never leave your device. Open source and deployed on Cloudflare.
- 100% client-side compression and conversion, no uploads
- Powered by ffmpeg.wasm running directly in the browser
- Free with no sign-up, deployed on Cloudflare Workers
- Open source: github.com/kumard3/vidshrink
- Video Processing
- ffmpeg.wasm
- WebAssembly
- Cloudflare Workers
- TypeScript
- React
- Client-side
- SEO