Gaurav Kumar.
Building full-stack systems and AI tooling.
Software engineer working at the seam between full-stack systems and applied AI — building AI agents with Claude Code, sales-workflow automation across n8n / Make / Clay, and on-device speech tooling (Whisper, XTTS-v2).
A note on the work.
I build the parts of products that have to be durable — backends that handle money, extensions that handle developer trust, and AI features that have to actually work when the stream stalls or the tool-call returns the wrong shape.
At PureCode AI I lead backend architecture for the enterprise service, ship tool-calls and a terminal manager into a VS Code extension, and tune prompts in the proxy that fronts our models. I also moved an entire payments stack from Paddle to Stripe — customers, active subscriptions, methods — without a minute of downtime.
Before that, I was a backend intern at Massive Mobility, where I tightened API contracts, cut payload sizes, and built the payments dashboard the team still uses to read transaction health.
The working stack.
Tools are means, not identities — but a clear inventory is honest. These are the ones I reach for first.
Languages
Frontend
Backend
Databases
AI / Agents
Automation & Sales
Cloud / Ops
Testing
Where the work happened.
PureCode AI
Backend and AI-tooling work across a VS Code extension, an enterprise service, and a payments stack — the kind of work that has to keep running while it changes underneath.
- Led backend architecture for PureCode Enterprise — Node.js, PostgreSQL, API design standards.
- Shipped a terminal manager and tool-calls (file I/O, command execution, terminal reads) into the PureCode VS Code extension.
- Built a team-shared rules feature so admins can organise and distribute response-quality rules.
- Authored CI/CD pipelines for App Service and Kubernetes deployments.
- Tuned prompts in the VS Code proxy service; measurable lift in model response quality.
- Migrated payments end-to-end from Paddle to Stripe — customers, active subscriptions, methods — zero downtime.
- Implemented LLM response streaming on LearnupAI to cut perceived latency.
- Automated the Outlook → LearnupAI content-injection pipeline, removing manual setup for QA.
Massive Mobility
Backend foundations for the Zecat and Bhamo apps, plus a payments dashboard that the team still reads daily.
- Hardened API contracts with frontend; reduced payload sizes and client-side latency.
- Reworked logging and exception handling for faster production triage.
- Built a real-time payments dashboard for transaction analytics.
- Deployed JWT-authenticated, validated, rate-limited APIs for Zecat and Bhamo.
Things I’ve shipped.
Hotinbox Sales Agent
An AI-driven sales workflow that wires Apollo, Clay, HubSpot and Instantly into a single automated lane — list building and enrichment on one end, multi-step outreach and CRM sync on the other. Built for sales teams who want pipeline, not plumbing.
- 01List building + enrichment pipelines across Apollo and Clay.
- 02n8n / Make-style automations syncing leads → CRM → outreach.
- 03Role-based access, campaign analytics, and response-rate dashboards.
PureCode AI · Terminal Manager
A terminal manager inside the PureCode VS Code extension — centralises command execution and terminal state so tool-calls can run, read back output, and stay coherent across sessions.
- 01Single source of truth for command execution across the extension.
- 02Tool-calls for file creation, shell execution, terminal reads.
- 03Reduced flakiness in AI-driven dev workflows.
VoiceType · Whisper-powered Dictation
A macOS menu-bar dictation app — push-to-talk hold-Control, on-device Whisper transcription, then injected straight into the focused field. A local-first clone of WhisperType / Wispr Flow.
- 01100% on-device transcription via pywhispercpp — no cloud round-trip.
- 02Voice-activity detection, hotkey capture, and overlay UI.
- 03Packaged as a signed .app with mic + accessibility entitlements.
Voxa · Multilingual TTS & Voice Cloning
A production-grade text-to-speech web app — 9 languages, voice cloning from a 6–30 s sample, auto-detected CUDA / Apple Silicon MPS / CPU backends. Model loaded once at startup for sub-second generation.
- 01XTTS-v2 voice cloning with custom speaker references.
- 02FastAPI routes for generate / history / download / clone-voice.
- 03Next.js App Router frontend with streaming playback + history.
Paddle → Stripe Migration
A full payment-gateway migration at PureCode — customers, active subscriptions, and payment methods moved with zero downtime for paying users.
- 01Mapped + migrated subscription state without a billing gap.
- 02Reconciliation pipeline to catch drift between providers.
- 03Cutover plan that kept charges flowing during the switch.
Let’s build something.
Open to roles, contracts, and the occasional weekend project — especially anything sitting at the line between product and AI. Reply times are generally within a day.