Currently building at Invsto

Built to
hold up
under load.

Backend & AI engineer. I build asynchronous, event-driven systems and the agentic AI platforms that run on top of them.

See the work →How I build
Pranshu Pandey, 3D avatarPranshu Pandey in focus mode// hover
Distributed systemsarchitecture
Ships to prodreliability
Agentic AIMCP · RAG
500+daily users on EI-LMS, self-hosted
15k+requests / day it serves
Jan 2026Full Stack & GenAI eng. at Invsto
▸ this site is playable — try it
Invsto · experienceEI-LMS · ERP / LMSHQ · life OSP.R.A.N. · reputation
↓ scroll
Experience

Invsto

Full Stack & GenAI Engineer

Jan 2026
→ present
8s → <1s
pipeline latency

My day job. I work across production repositories — AI applications, backend services, developer tooling, and the infrastructure underneath them.

What I reach for
FastAPIAsyncIORabbitMQRedisARQPostgreSQLMCPAzure AI FoundryRAGLangChainDockerNGINXAzureGitHub Actions
01

AI & agents

Systems that use models, and the guardrails around them.

  • —Built agentic AI platforms across web and CLI using MCP, Azure AI Foundry, and modular RAG architectures.
  • —Developed AI assistants for feature automation using retrieval systems and custom MCP tools.
  • —Built internal agents supporting developer productivity, UAT, and go-to-market workflows through tool orchestration.
  • —Developed secure execution environments for user-defined workflows, with strict runtime isolation and controlled tool access.
02

Backend & distributed systems

The asynchronous plumbing everything else stands on.

  • —Designed asynchronous, event-driven backends with FastAPI, AsyncIO, RabbitMQ, Redis, and ARQ workers.
  • —Re-architected a latency-critical processing pipeline onto distributed workers with Redis coordination and message queues — 8 seconds to sub-second.
  • —Built reusable infrastructure for retrieval, workflow orchestration, task scheduling, tool execution, and agent lifecycle management.
03

Platform & infrastructure

Owning the boxes, not just the code on them.

  • —Owned and contributed to multiple production repositories across AI applications, backend services, developer tooling, and platform infrastructure.
  • —Managed production infrastructure: Linux servers, SSH, NGINX reverse proxies, SSL/TLS, Azure deployments, container registries, and GitHub Actions CI/CD.
▸play with the latency re-architecture
Latency lab · interactive
8.0ssequential
0 completed · 0 queued
ingress → queue → workers → dbillustrative · drag the slider
02

EI-LMS

The department, running on rails.

A full-scale ERP and learning management system for the Electronics & Instrumentation department — attendance, coursework, testing and reporting in one place. Self-hosted and kept alive in production.

  • —QR-based attendance, assignments, forms, notes distribution and simulator integrations.
  • —An online testing platform with automated evaluation, analytics dashboards and email-delivered results.
  • —Automated the department’s PDF and Excel reporting workflows.
  • —Self-hosted behind NGINX with PM2.
500+daily users
15,000+requests / day
ReactExpressPostgreSQLPrismaZustandNGINXPM2
Live ↗Client ↗Server ↗
In production
03

HQ

A command center for one person.

A single-user command center that runs on my own machine against a local SQLite file — job hunt, work, projects, reflections, notes, people, gym, wellbeing. The web app is only half of it: a native macOS shell wraps it in the executive-function layer a browser tab cannot provide, and an Android build is next.

  • —A floating always-on-top beacon: a wall clock that becomes the running timer the moment a focus session starts. Local state, so it never depends on the server.
  • —Global hotkeys for capture — start a session, catch a stray task, log a rabbit hole mid-focus, or collapse everything to one thing when overwhelmed.
  • —Starting a session flips macOS into Do Not Disturb via Shortcuts and opens the tied workspace; a context shield notices when you slip into a time-sink app and nudges once.
  • —A menu-bar panel with what you are doing right now, what is due next, and what you set aside.
  • —Ships its own MCP server (~45 tools, stdio + streamable HTTP) so Claude, Cursor or ChatGPT can operate it.
  • —A dialectic journal, a polymorphic mention graph, wellbeing targets derived from a profile, and an XP system that rewards effort rather than outcomes.
Localsqlite, on device
~45MCP tools
macOSnative · Android next
Next.js 16React 19ElectronTailwind v4PrismaSQLiteTipTapMCP
Source ↗No hosted demo by design — it is single-user and keeps your data on your own machine. Clone it and run it yourself.
local · sqlite
04

P.R.A.N.

Reputation. Engineered.

An AI-powered online reputation management platform. It scrapes and watches what the internet is saying, runs sentiment analysis over it, and turns that into brand monitoring and audience-growth recommendations — presented as a HUD console.

  • —Distributed scraping and sentiment-analysis pipelines.
  • —RAG-based analysis workflows feeding growth recommendations and brand monitoring.
  • —Modular backend services and scalable dashboards with clear API boundaries.
FastAPINext.jsPlaywrightRedisRAG
Live ↗Source ↗
ACCESS_GRANTED
Control room

This site is a system.
You can operate it.

I build things people run, not things people look at. So rather than describe that, the site lets you run it — and break it.

Your session
achievements
0/14
CV artifacts
0/7
best wave
—
queue drained
—
best MTTR
—
⌘K anywhere · ` for the shell
Writing

Notes from the queue.

Mostly about the unglamorous half of the job — latency, retrieval, and the boundaries you put around things that can act on their own.

  1. 01The spinner was lyingEight seconds of "loading" for work that nobody was actually waiting on.Systems12 Jun 20262 min
  2. 02Five decisions in a trench coatEveryone calls it "retrieval", as if it were one thing you could get right once.AI2 May 20262 min
  3. 03Nobody owns the middleThe queue between two services is not plumbing. It is the only contract they share.Architecture8 Apr 20262 min
  4. 04The queue that ate its workerAn unbounded queue is not resilience. It is an outage with excellent manners.Systems15 Mar 20262 min
  5. 05An unattended shell with good grammarThe moment you hand an agent real tools, you have shipped something with opinions and access.AI20 Feb 20262 min
Approach

How I build.

I'm Pranshu — a backend and AI engineer who likes the parts of the job that don't demo well: queues, workers, retrieval, the plumbing behind the screen. Currently doing that at Invsto.

  1. 01

    Do the work asynchronously

    Most things do not need to happen while someone waits. Queues, workers and event-driven boundaries are how a system stays responsive — and how an eight-second pipeline becomes sub-second.

  2. 02

    It only counts in production

    A system is real when people depend on it. That means self-hosting it, watching it at 15k requests a day, and owning the servers, certificates and pipelines underneath.

  3. 03

    Give the AI real tools, and a sandbox

    Agents are only as good as what they can reach. I build retrieval, tool execution and lifecycle plumbing — with strict runtime isolation, because an agent with real tools needs real boundaries.

Contact

Building something that
needs to hold up?

I'm open to interesting problems — distributed backends, agentic AI systems, or the fast-and-reliable end of product engineering.

Get in touch →Résumé ↗
GitHub ↗LinkedIn ↗X / @notoriouspran ↗Indore, India
system nominalup 00:0060fps0 nodes
control room ↑