Skip to content
AI agents · RAG · MCP · .NET · Kafka

I build AI agentsand backends at scale.

Software engineer at Falcon Autotech. I build the C#/.NET services that route parcels for sortation systems in India, the UK and Saudi Arabia, and the Python agents and RAG pipelines that run alongside them.

Explore projectsOpen resumeTry the RAG Visualizer
Open to AI and backend engineering rolesNoida, India

PRODUCTION PATH

.NET services → messaging → equipment

requests sustained per hour
1.5M+
parcels coordinated per hour
50k+
countries with live deployments
3
years in production engineering
2.5

PROJECTS

Three labs and two builds.

Three labs you can use in the browser, an AI incident-triage assistant, and an architecture case study of a notification service.

3 passages · illustration

Query

How does grounding reduce hallucinations?

Retrieved contextScore
RAG grounding0.98
Source attribution0.93
Fallback policy0.81
AI / 01RETRIEVAL SYSTEMSSTART HERE

RAG Visualizer

A browser lab for inspecting how a query becomes ranked context and then an answer — on your own documents.

384

dimensions in local mode

  • Bring your own PDF, DOCX, TXT or pasted text, then chunk, embed and search it in cloud mode or fully in the browser
  • Real 384-dimension embeddings from all-MiniLM-L6-v2 running on WebGPU/WASM via transformers.js in a worker
  • Answers from a cloud model by default, or from SmolLM2-360M running locally on WebGPU
Run a retrieval trace
NEXT.JSTYPESCRIPTTRANSFORMERS.JSWEBGPU
rate limiterquota · illustration
3requests left
algorithm
token bucket
refill
2 / sec
responses
200 / 429
EDGE / 02TRAFFIC CONTROLINTERACTIVE LAB

LimitLine

A playable rate limiter: a live quota you can watch, requests you fire one at a time, and the exact reason each one came back 200 or 429.

3

algorithms under test

  • A real-time gauge of how many requests you can still make, and when the next one unlocks
  • One click sends one request along a visible pipeline, decided the moment it reaches the limiter
Spend the quota
NEXT.JSTYPESCRIPTSVGREQUESTANIMATIONFRAME
illustration

Incident

Example: error rate rising after a deploy

KafkaRetrieveTriage
EvidenceScore
Runbook match0.91
Similar past incident0.86
Service log window0.78
AI / 03INCIDENT RESPONSENO PUBLIC DEMO

AI Incident Triage

An assistant for production incidents. It reads logs and service events from Kafka, finds the matching runbooks and past incidents with RAG, and bases its triage and root-cause suggestions on them.

3

evaluation axes: retrieval, latency, groundedness

  • Evaluated on retrieval quality, latency and groundedness
  • Built on .NET with Kafka for events and Redis for state, packaged with Docker
AGENTIC RAG.NETKAFKAREDIS
cache-aside · illustration
user:1hit
18s
product:2hit
09s
order:1evict
LRU
cart:4miss
DB
ClientCacheDatabase
DATA / 04CACHING SYSTEMSINTERACTIVE LAB

CacheFlow

An interactive cache-aside simulator with visible hits, misses, expiry, and eviction.

5

cache events exposed

  • Animated request flows between client, cache node, and database
  • Cache-aside strategy with hits, misses, invalidation, TTL countdowns, and LRU eviction
Explore the cache
NEXT.JSTYPESCRIPTTAILWIND CSSSVG
7 channels

One publish contract

IntakeQueueWorkers
urgentpriority
standardqueued
bulkthrottled

Channel adapters

EMAILSMSPUSHWEBHOOKSLACKTEAMSWHATSAPP
MESSAGING / 05NOTIFICATION INFRASTRUCTURECASE STUDY

Distributed Notification Service

A high-throughput, queue-based notification service designed to dispatch more than one million notifications per day across seven delivery channels.

1M+

notifications/day design target

  • One publish contract fans out to email, SMS, push, webhook, Slack, Teams, and WhatsApp adapters
  • Priority queues, scheduling, and rate limits keep urgent traffic moving without overwhelming downstream providers
Read the case study
QUEUE-BASED DELIVERYPRIORITY QUEUESEXPONENTIAL BACKOFFDLQ

Or play it: a 90-second 3D run where you are the query, and the answer about me is built from whatever you retrieve. There is also a warehouse rover to drive.

Play GroundedDrive the warehouse roverGitHub profile

EXPERIENCE

Where I've worked.

Software engineer at Falcon Autotech since May 2024. Before that, a full-stack internship at Inxee Systems.

GENERATIVE AI AND DISTRIBUTED SYSTEMS

Software Engineer, AI & Backend

Falcon Autotech

AI features, and sortation services in three countries

I build AI agents, RAG pipelines and LLM integrations in Python, and the C#/.NET services between warehouse software and sortation machines.

AGENTS AND MULTI-AGENT WORKFLOWS
Design, build and deploy AI agents and multi-agent workflows with LangGraph, LangChain and CrewAI, integrated with internal services, tools and APIs.
RAG PIPELINES
Implement RAG pipelines for enterprise knowledge retrieval and question answering: ingestion, chunking, embeddings, vector search, re-ranking and grounded generation.
LLMS VIA MCP AND TOOL CALLING
Integrate LLMs (GPT, Claude, Gemini, Llama) with enterprise applications through MCP and tool calling, and build Python (FastAPI) services that serve AI features alongside the .NET platform.
EVALUATION IN PRODUCTION
Tune prompts, agent workflows, latency and cost, and run evaluations that track answer quality, groundedness and reliability in production.
SORTATION SERVICES
Build and operate event-driven C#/.NET microservices between warehouse software and sortation systems in India, the UK and Saudi Arabia: 1.5M+ requests and 50,000+ parcels an hour over Kafka, RabbitMQ, Redis and MySQL, with UDP/TCP sockets to the hardware.
20% MORE EFFICIENT · 5/5 RATING
Improved system efficiency by 20% and cut bug-related downtime by 15% through performance tuning, observability and root-cause analysis; earned a 5/5 performance rating.
WORKING STACKPythonFastAPILangGraphLangChainCrewAIMCP.NETC#KafkaRabbitMQRedisMySQL
INVENTORY OPERATIONSINTERNSHIP

Web Application Developer

Inxee Systems Pvt Ltd

End-to-end product delivery

Shipped a custom inventory system across the interface, API, and persistence layers.

25% FEWER DISCREPANCIES
Built inventory workflows that reduced stock discrepancies by 25%.
30% BETTER ORDER ACCURACY
Improved order accuracy by 30% through clearer stock and fulfilment state.
25% FASTER PAGE LOADS
Improved React interface performance, reducing page-load time by 25%.
FULL-STACK DELIVERY
Implemented Node.js and Express services with MongoDB persistence.
WORKING STACKReactNode.jsExpressMongoDB

SKILLS

Tools and stack.

Agent and RAG tooling first, then the services, queues and data stores I run in production.

CERTIFICATIONS

  • Google Cloud Fundamentals: Core Infrastructure
  • Essential Google Cloud Infrastructure: Foundation
  • Architecting with Google Kubernetes Engine: Foundations

AI agents

Agents and multi-agent workflows wired to real tools, services and APIs.

  • LangGraph
  • LangChain
  • CrewAI
  • AutoGen
  • Semantic Kernel
  • MCP
  • Tool calling

LLMs and retrieval

RAG from ingestion to grounded answers, measured with LLM evaluations.

  • GPT
  • Claude
  • Gemini
  • Llama
  • RAG
  • Re-ranking
  • Pinecone
  • FAISS
  • Chroma
  • Weaviate
  • Milvus
  • LLM evaluation
  • Fine-tuning

Services and APIs

C#/.NET and Python services in production, and the REST APIs in front of them.

  • C#
  • .NET Core
  • ASP.NET Web API
  • Python
  • FastAPI
  • TypeScript
  • Node.js
  • Go
  • REST

Messaging and data

Kafka and RabbitMQ between services; Redis, MySQL, PostgreSQL and MongoDB for state.

  • Kafka
  • RabbitMQ
  • Redis
  • MySQL
  • PostgreSQL
  • MongoDB

Real-time integration

UDP and TCP socket code that sends commands to sortation equipment.

  • UDP
  • TCP

Cloud and delivery

Containers, pipelines and observability for production services.

  • AWS
  • Azure
  • Google Cloud
  • Docker
  • CI/CD
  • MLOps
  • Observability
  • Linux

Frontend

Enough frontend depth to make backend state inspectable, useful, and testable.

  • TypeScript
  • React
  • Tailwind CSS
  • SVG
  • Vitest

ABOUT

What I do day to day.

At Falcon Autotech I build the C#/.NET services that tell sortation equipment where each parcel goes, over Kafka, RabbitMQ, Redis and MySQL. Alongside them I build Python (FastAPI) services for AI features: agents, RAG for enterprise knowledge, and LLMs connected to internal tools through MCP, with evaluations for answer quality and groundedness.

CONTACT

Email me aboutAI or backend roles.

Email is the quickest way to reach me.

[email protected]

Resume

PDF, one page

GitHub

Profile and activity

LinkedIn

Work history and messages

LeetCode

Problem-solving record

Open to AI and backend engineering roles

Developed by Abhishek Suman with