Skip to content
PG
Prince Gupta, smiling, in a white hoodie and glasses

~/nastyrunner13

Prince Gupta

AI Engineer · multi-agent systems, tool use & evals

Building agents @ Tata CommunicationsBengaluru, India

I build multi-agent systems that hold up in production: LangGraph orchestration, tool calling and memory that survive real enterprise workflows, and the eval pipelines that keep them honest.

POST/v1/whoami200 · 60 tok
{"role": "AI Engineer","company": "Tata Communications","focus": ["multi-agent systems", "tool calling", "memory", "evals"],"stack": ["Python", "TypeScript", "LangGraph", "FastAPI"],"location": "Bengaluru, India","github": "NastyRunner13"}
Email me Résumé

About

~1 min read

Hello.

I build agents for real customers. At Tata Communications I designed and shipped an enterprise multi-agent platform on LangGraph and FastAPI, with hierarchical delegation and ReAct orchestration. It serves 500+ enterprise customers across billing, ServiceNow ticketing and support.

Reliability is the job. Agents fail in boring ways: a flaky API, lost context, a tool call that times out. I design for that with node-level retries, fault-tolerant clients, summarized long-term memory and eval pipelines that hold accuracy and latency thresholds.

After hours, I build in the open. Desktop tooling for AI coding agents, multi-agent research systems and open-source takes on Claude Code and Cowork. Most of it lives on my GitHub.

building
TeamBot, an open-source office for AI agents
exploring
earned autonomy for desktop agents in AgentOS
studied
B.Tech CSE (AI & ML), VIT Bhopal, 2025

Experience

3 roles

  1. Mar 2025 — Present

    Bengaluru, India

    Building the enterprise multi-agent platform behind billing, ServiceNow ticketing and support workflows for 500+ enterprise customers.

    multi-agent platform

    • Designed and deployed a multi-agent platform on LangGraph and FastAPI with hierarchical delegation and ReAct-based orchestration, serving 500+ enterprise customers.
    • Architected agentic workflows with custom state management, tool-calling interfaces and summarized long-term memory, so multi-step tasks run reliably across complex enterprise use cases.

    reliability & evals

    • Built eval pipelines and resilience layers with node-level retries in LangGraph and Tenacity-based fault-tolerant API calls, significantly reducing workflow failures while holding accuracy and latency thresholds.

    integration

    • Shipped production APIs with FastAPI and Docker that connect the AI platform to customers' live infrastructure, including legacy APIs and support systems.
    • Python
    • LangGraph
    • FastAPI
    • ReAct
    • Tenacity
    • Docker
    • ServiceNow

  2. Jun 2024 — Aug 2024

    Pune, India

    Built Transcripto, a Streamlit app for transcription, summaries and Q&A over videos, PDFs and images.

    multimodal ai

    • Integrated Wav2Vec2 for speech-to-text and EasyOCR to pull text out of images.
    • Used PyPDF2 and MongoDB for PDF text extraction and storage.
    • Python
    • Streamlit
    • Wav2Vec2
    • EasyOCR
    • MongoDB

  3. Video Production Manager

    Photon Ecademy

    Feb 2023 — Dec 2023

    Remote

    Led a team of five in-house video editors producing educational content on tight deadlines.

    leadership

    • Delivered 50+ content pieces by reviewing scripts and working closely with creators.

education

Vellore Institute of Technology, Bhopal

B.Tech in Computer Science, specialization in AI & ML

Aug 2021 — May 2025

CGPA 8.24

Selected work

3 case studies

NexusTrade

Multi-agent trading system · 2026

Source

problemOne prompt can't weigh market data, sentiment, news and fundamentals the way a trading desk does, or argue with itself before committing.

  • A LangGraph DAG of 13 specialized agents: four analysts run in parallel (market, sentiment, news, fundamentals), then a bull/bear research debate and a five-agent risk committee.
  • Decisions stream over FastAPI WebSockets to a Next.js dashboard while the agents work.
  • ChromaDB memory of past trades, EMA-weighted confidence tracking, Kelly-capped position sizing and a rate-limited Groq client with backoff, so one pipeline handles stocks (yfinance) and crypto (CCXT).
spec
agents
13 specialized
flow
4 parallel analysts → bull/bear debate → 5-agent risk committee
memory
ChromaDB over past trades
sizing
Kelly-capped, EMA-weighted confidence
  • Python
  • LangGraph
  • LangChain
  • FastAPI
  • ChromaDB
  • Next.js
  • Groq

AgentNotch

Desktop observability for AI coding agents · 2026

Source

problemWith several AI coding agents running at once, it's hard to tell what each one is doing, waiting on or costing.

  • A cross-platform Electron app that watches Claude Code, Codex, Cursor, Grok Build, OpenCode and Antigravity in real time through log tailing, SQLite reads and process watchers.
  • Local-first by design: sandboxed renderer, secure IPC through contextBridge, read-only file inspection and zero telemetry.
  • A system-tray notch UI with session dispatch, on-device usage and cost analytics, and approve/deny for Claude permission prompts right from the tray.
spec
watches
Claude Code, Codex, Cursor, Grok Build, OpenCode, Antigravity
platforms
Windows, macOS, Linux (NSIS, DMG, AppImage)
telemetry
none, everything stays on your machine
ci
GitHub Actions on 3 OSes × Node 20/22
  • JavaScript
  • Electron
  • Node.js
  • Chokidar
  • GitHub Actions
  • electron-builder

TeamBot

Self-hosted office for AI agents · 2026

Source

problemGiving agents real work means giving them real tools, and most setups offer little isolation, oversight or cost control.

  • Named agents, each with its own model (any tool-calling model on OpenRouter) and its own isolated Linux computer with a terminal and a Chromium browser you can watch live and take over.
  • Every tool call passes a policy check (allow, review, ask, deny, hand off), with daily and monthly budgets per agent.
  • Prompt-injection defence for web pages, files and MCP results, secrets filled in only at execution time, and durable runs that survive restarts.
spec
isolation
a Linux container per agent
governance
allow · review · ask · deny · hand off
budgets
dollar and token caps per agent
models
any tool-calling model via OpenRouter
  • TypeScript
  • Fastify
  • WebSockets
  • Docker
  • MCP
  • React
  • noVNC

How I build

one agent run, traced

Most LLM features I ship look roughly like this under the hood. Hover or tap a span to see why it's there.

trace a91f3c079 spans · 3.12s
supervisor.routellm→ billing agent

Hierarchical delegation: a supervisor reads the request and hands it to the right specialist agent.

  • chain
  • llm
  • retrieval
  • tool
  • guardrail
  • eval

  • Delegate, don't overload

    A supervisor routes work to focused specialists instead of one prompt trying to do everything.

  • Retry at the node

    Failures are expected. Each node retries on its own, so one flaky API doesn't sink the run.

  • Memory as summaries

    Long-term memory is summarized, so agents remember the past without carrying all of it.

  • Evals gate releases

    Accuracy and latency thresholds are checked by a pipeline, not by eyeballing outputs.

More things I've shipped

16 projects

2026
2025
2024

Activity

667 contributions · last 12 months

@NastyRunner13 on GitHub LessMore

Stack

39 tools I reach for

languages
  • Python
  • TypeScript
  • JavaScript
  • C++
agents
  • LangGraph
  • LangChain
  • LlamaIndex
  • MCP
  • ReAct
  • Multi-agent orchestration
models
  • Claude
  • GPT
  • Gemini
  • Llama
  • Groq
  • Ollama
  • OpenRouter
retrieval & memory
  • RAG
  • ChromaDB
  • MongoDB
  • Long-term memory
evals & reliability
  • Eval pipelines
  • Tenacity
  • Node-level retries
ml
  • PyTorch
  • TensorFlow
  • Hugging Face
  • scikit-learn
  • LoRA / QLoRA
backend & infra
  • FastAPI
  • Fastify
  • WebSockets
  • Docker
  • OAuth 2.0
  • GitHub Actions
apps
  • Next.js
  • React
  • Electron
  • Streamlit

Follow along

AI content on Instagram

Want more AI in your feed?

I post about AI on Instagram: agent builds, tools worth trying, and what I learn shipping LLM products.

Follow @princegupta.ai

Contact

usually replies within 24h

Building something with agents?

I like talking about multi-agent systems, evals and the unglamorous parts of shipping LLM products. Send a note and I'll get back to you.

$ compose hire.me
to:princegupta1586@gmail.com

Replies within 24h

Or find me elsewhere