🟢Online· live

SiddarthaNallella

AI-Native Full-Stack Engineer. I build AI products people actually use — for real companies, real users, and real deadlines.

See the work
github.com/harshasiddartha
0M
requests a second, on a web server I wrote from scratch
0+
bookings taken by a platform I built
0
fixes accepted into major open-source projects
0
projects built and documented

●At the keyboard and replying fast.see what I'm up to

LLVMApache DataFusionDockerHuggingFaceRayKubernetesMCP SDKMicrosoft AspireORASDeepEvalpromptfooValePyMCpandasThunderbirdFirebaseOrchardCoreMDNastropyQwik UIDrizzlePHP-CS-FixerLLVMApache DataFusionDockerHuggingFaceRayKubernetesMCP SDKMicrosoft AspireORASDeepEvalpromptfooValePyMCpandasThunderbirdFirebaseOrchardCoreMDNastropyQwik UIDrizzlePHP-CS-Fixer
At a glance

The short version.

What I am doing right now, what I have built, and what has been accepted into other people's projects — on one screen.

right now

🟢 Online

At the keyboard and replying fast.

Hyderabad, India open to roles

Peak throughput

0.0M

requests a second, on a web server I wrote from scratch

My local time

--:--:--

IST · UTC+5:30

100%

0

tests across the systems projects

Languages

  • JavaScript96
  • CSS83
  • TypeScript73
  • HTML26
  • Python21

Repositories

0

202420252026

Upstream contributions

0pull requests

14 merged · 14 in review

Flagship buildRaft Distributed KV StoreA database that survives servers crashing.600failure scenarios simulated
Latest postTesting Raft with a deterministic simulatorWhy I injected the clock, the RNG and the transport into my C++20 Raft implementation, and what 600 seeded chaos steps actually check.2 min read

Shipped & documented

0

projects on this site, each with a public repo

01 — About

I build software that companies pay for.

Two years shipping AI products to real users, plus the low-level engineering most people leave to a library. Below is what that means in practice.

I ship products, not prototypes

Two years building AI features that companies actually launched — an assistant inside a health-tech platform, automated document workflows at Quanta Global, a prototype network agent at Ericsson, and a freelance booking platform that handled over 12,000 orders.

I build the hard parts myself

When something has to be fast or has to stay up, I write it rather than reach for a library. A database that survives servers crashing. A web server handling 2.8 million requests a second. An AI model runner that needs no graphics card.

I make sure it does not break

Every project ships with its own tests — 454 of them across my systems work. I find real bugs in other people's code too: 28 of my fixes have been accepted by the teams behind LLVM, Docker, HuggingFace, Ray and Kubernetes.

Education

IIITDM Kancheepuram

B.Tech, Computer Science & Engineering · 2022 – 2026

Right now

SDE I @ Xenotix Labs

Building ride-hailing, food-subscription and community platforms that are live and taking payments.

Looking for

My next role

AI products, backend engineering, or full-stack. Remote or Hyderabad — happy to walk through any project here.

Right now

What I am working on this month.

I update this myself whenever it changes. The status light at the top of the page comes from the same place.

🟢Online· live

Focus

Shipping ride-hailing and subscription-commerce backends at Xenotix Labs, and keeping a freelance pipeline alive on the side.

Learning

RL environments, evals and benchmarks — how frontier labs build and grade agent tasks.

Building

game changing for aspirants

the full now page
02 — By the numbers

Two years of work, counted.

Pulled straight from my GitHub account. Nothing rounded up.

Languages across every repository

JavaScript
96
CSS
83
TypeScript
73
HTML
26
Python
21
Shell
10
C++
7
PL/pgSQL
6
Kotlin / Swift
5
CMake
4

Repositories created

0

24
2024
63
2025
43
2026

75 public · 55 private · excludes 49 contribution forks

Tests written

0

Every systems project ships its own suite.

Where those tests live

  • 300
    DSA & systems
  • 47
    Coding agent
  • 35
    Inference server
  • 30
    HTTP server
  • 23
    Vector DB
  • 19
    Task queue

Open-source reach

0pull requests

across 25 projects in 24 orgs

0merged0in review
03 — Experience

Five roles. Two years. Real users waiting on it.

  1. Aug 2026 – Present

    Remote

    current

    Xenotix Labs

    Software Development Engineer I
    • Architected a ride-hailing and parcel delivery backend as a two-process system — API gateway and BullMQ workers scale independently, so job throughput never blocks request latency.
    • Integrated a payment gateway end to end: signed callback verification per signing mode, platform-specific callback keys, and a runtime test/live toggle that removed the redeploy step.
    • Delivered a subscription commerce platform against a client-signed spec, encoding recurring cycles, pause limits, wallet gates and tiered delivery fees as enforced data-model invariants.
    • Shipped a multi-tenant community management suite and remediated two cross-tenant data-exposure defects in the authorization layer.
    Node.js 24TypeScriptExpress 5MongoDBRedisBullMQSocket.IONext.jsFlutterReact Native

  2. Mar 2026 – Aug 2026

    Bengaluru, India

    Ericsson

    Research Intern, Compute & Software
    • Engineered a prototype AI agent for 5G networks with FastAPI, LangGraph and FastMCP — natural-language-driven user session creation and modification from application-layer intent.
    • Benchmarked agent workflow configurations to cut end-to-end response latency.
    • Researched LLM serving, AI Agent Operating Systems, dynamic scheduling and agentic memory (A-Mem), folding disaggregated-serving strategies into prototype designs.
    FastAPILangGraphFastMCPLLM servingAgent memory

  3. Feb 2025 – Dec 2025

    Remote

    Quanta Global

    Software Developer Intern
    • Built and deployed production LLM workflows with LangChain, LangGraph, Google GenAI SDK and Vercel AI SDK, automating manual processing through reasoning pipelines.
    • Designed enterprise RAG systems on Pinecone, Astra DB and MongoDB that raised semantic retrieval precision over keyword search.
    • Optimized REST APIs across FastAPI, Express, PostgreSQL and Prisma; containerized AI microservices with Docker.
    Next.jsFastAPISupabaseLangChainPineconeAstra DBDocker

  4. Nov 2024 – Present

    Remote

    current

    Freelance

    Full-Stack Developer
    • Engineered an event platform that handled 12,000+ bookings — secure payments, automated SMS, a full admin dashboard and an enquiry chatbot.
    • Delivered real-estate, EdTech and retail platforms with automated email workflows and conversational enquiry capture, plus custom n8n backend automations.
    Next.jsSupabasePostgreSQLn8nPayments

  5. Oct 2024 – Dec 2024

    Remote

    Clinics on Cloud

    Health-Tech Platform Revamp
    • Migrated the core health-tech platform's legacy interfaces to a responsive React + Tailwind application.
    • Built metrics-driven internal dashboards and shipped a Python/REST AI assistant automating internal workflows and patient queries.
    ReactTailwindPythonPostgreSQL
04 — Selected work

Things I have built.

Products people use, and the machinery underneath them. Every card links to the code — nothing here is a slide deck.

systemsview repo

Raft Distributed KV Store

A database that survives servers crashing.

Storage spread across several machines that keeps working when one dies or the network splits in half. I wrote the agreement algorithm from scratch in C++ and proved it correct by simulating 600 different failure scenarios — each one replayable on demand from its seed.

600failure scenarios simulated
C++20RaftgRPCWALASan/TSanPorcupine
systemsview repo

High-Throughput HTTP Server

A web server handling 2.8 million requests a second.

Written directly against the operating system with no framework underneath. Every CPU core runs its own independent copy and shares nothing with the others, so adding cores adds speed instead of contention.

2.8Mrequests a second, on one core
C++20epollLinuxZero-copy
systemsview repo

CPU LLM Inference Server

Runs AI language models on an ordinary laptop CPU.

No graphics card and no existing AI library. I wrote the model loader, the maths kernels and the text tokeniser myself. It answers on the same API as OpenAI, so tools built for ChatGPT work against it unchanged.

35tests, each checked against a reference
C++20AVX2GGUFTransformersSSE
systemsview repo

HNSW Vector Database

The search index behind AI 'find similar' features.

Holds millions of items and returns the closest matches in milliseconds, built straight from the original research paper. A damaged data file is refused outright rather than quietly misread.

23tests, including a corruption check
C++20HNSWSIMDmmap
aiview repo

Agentic Coding Agent

An AI agent that fixes bugs and checks its own work.

Point it at a codebase and a failing test. It reads the code, makes an edit, runs the tests, and tries again when it is wrong. All of it happens inside a five-layer sandbox — I attacked that sandbox ten different ways and it blocked every one.

100%of attempted sandbox escapes blocked
LangGraphQdranttree-sitterSandboxingLangfuse
aiview repo

Deep Research Agent

Researches a question and cites every single sentence.

Breaks a question into parts, reads sources for each, then grades its own coverage and goes back for more where it is thin. Every sentence in the final report carries a citation, checked automatically rather than trusted.

100%of sentences carry a source
LangGraphHybrid RAGSSEGuardrails
aiview repo

AstraRAG Enterprise

AI search across a company's own documents.

Many customers share one system and none can ever see another's data — that separation is enforced in three different databases at the same time. If one AI provider hits its usage limit, it switches to another without dropping the request.

3databases enforcing customer separation
Graph RAGNeo4jQdrantPostgreSQLNext.js 15
systemsview repo

Distributed Task Queue

Makes sure background jobs actually finish.

Work is picked up again when a machine dies mid-task, scheduled jobs run on time, repeated failures are quarantined instead of vanishing, and submitting the same job twice never runs it twice.

19tests, in under three seconds
PythonFastAPIPrometheus
aiview repo

E2E MLOps Pipeline

Tracks an AI model from training to production.

Versions the data, records every experiment, and only promotes a new model if it beats the current one during automated checks. Every outside service has a local stand-in, so the whole pipeline runs on a laptop.

PythonMLOpsCIFAISS
aiview repo

ResumeIQ

Scores a CV against a job description.

Upload a résumé and a job post. It pulls out contact details, skills and experience, scores how well the two match, and lists the keywords that are missing. Runs entirely on AWS with no server to manage.

AWS LambdaBedrockTextractS3
toolingview repo

LeetUI

A coding-practice app that lives in your terminal.

Published on PyPI. Browse problems, write and run solutions side by side, and track how much of a study list you have covered — without ever leaving the keyboard.

PythonTextualPyPI
toolingview repo

dictapp

Voice dictation for Mac that never touches the internet.

Hold a key, speak, and the words appear in whatever you were typing in. No account, no API key, and no audio ever leaves the machine.

Electronwhisper.cppmacOS
Bonus — 13 rebuilds, 300 tests

I rebuilt the building blocks hiding inside databases, caches and search engines.

Thirteen of them, written from scratch — each with its own tests and a measurement against the simpler approach it replaces.

LRU / LFU cache

mirrors Redis maxmemory

01

Rate limiter

mirrors Stripe / Cloudflare

02

Bloom · CMS · HyperLogLog

mirrors Redis PFCOUNT

03

Autocomplete (Trie, BK-tree)

mirrors Search suggest

04

Search engine (BM25)

mirrors Lucene

05

Diff (Myers)

mirrors git diff

06

LSM key-value store

mirrors RocksDB

07

B-tree index

mirrors PostgreSQL

08

Consistent hashing

mirrors DynamoDB

09

Job scheduler (Kahn)

mirrors Airflow

10

Route planner (A*)

mirrors Maps routing

11

Mini git (Merkle)

mirrors git

12

Compressor (Huffman, LZ77)

mirrors gzip

13
05 — Under the hood

Proof it actually works.

Real output from my own test suites. Pick a tab and watch one run — every number on this site comes from a test like these.

❯
06 — Open source

My fixes run inside tools you already use.

28 contributions accepted or under review at LLVM, Docker, HuggingFace, Ray, Kubernetes and more. Real bugs — a crash, a race condition, silently wrong results — not spelling corrections.

28
pull requests
14
merged
14
under review
07 — Writing

I can explain what I build.

all posts
20 August 2026 · 2 min

Testing Raft with a deterministic simulator

Why I injected the clock, the RNG and the transport into my C++20 Raft implementation, and what 600 seeded chaos steps actually check.

distributed-systemsraftc++20testing
14 July 2026 · 1 min

2.8M requests/sec on one core: what a share-nothing reactor buys you

Notes from writing an HTTP/1.1 server in C++20 against raw Linux syscalls — epoll, SO_REUSEPORT, sendfile — and a parser that never allocates.

c++20linuxnetworkingperformance
30 June 2026 · 1 min

Running Llama on a CPU with nothing but C++20

A single-request inference server with no GPU, no ggml and no llama.cpp: GGUF loading, quantized matmuls, a BPE tokenizer and an OpenAI-compatible streaming API.

llminferencec++20simd
papers, docs and tools I keep coming back to
08 — How I work

The rules I hold myself to.

If something can fail, I test that it fails safely.

— Distributed database

Bad configuration should stop a deploy, not wake someone at 3am.

— Ride-hailing backend

A corrupted file gets refused, never quietly misread.

— Vector database

Every bug should be reproducible on demand, never 'sometimes'.

— Simulation testing

Swapping out a service should be a config change, not a rewrite.

— ML pipeline

A slow background job must never slow down someone's request.

— Two-process architecture

Trust a heartbeat, not the network's word for it.

— Job scheduler

Prices live in config, so a business change isn't an engineering change.

— Subscription platform
09 — Stack

What I actually work with.

Everything listed here is something I have shipped with, not something I have only read about.

C++20PythonTypeScriptRustGoLangGraphRaftHNSWepollAVX2gRPCQdrantNeo4jPostgreSQLRedisMongoDBFastAPIExpressNext.jsReactTailwindThree.jsFlutterReact NativeDockerAWSBedrockGeminiGroqMCPLangfuseRAGASBullMQRabbitMQPrometheusSupabasePrismaDrizzleSnowflakeFAISSPineconeLLVMKuberneteswhisper.cppGGUFSIMDWALPorcupine

Languages

C++20CPythonTypeScriptJavaScriptGoRustSQL / PL-pgSQLDartTCL

AI / LLM

LangGraphLangChainMCPGraph RAGHybrid dense + BM25RRF fusiontree-sitterStructured outputsTool callingLangfuseRAGASRed-team evalsBedrockGeminiGroqVercel AI SDKwhisper.cpp

Systems & distributed

RaftWAL / fsync orderingConsistent hashingDeterministic simulationPorcupinegRPCepoll ETSO_REUSEPORTsendfile(2)AVX2 + FMAmmapASan / TSanCMake

Backend & data

FastAPIExpress 5Node.js 24Socket.IOBullMQRabbitMQRedisPostgreSQLMongoDBNeo4jQdrantPineconeSupabasePrismaDrizzleSnowflakeFAISS

Frontend & mobile

Next.js 16React 19Server ComponentsTailwind v4shadcn/uiRadix UITanStack QueryFramer MotionThree.jsFlutterReact NativeExpo

Cloud & platform

AWS LambdaAPI GatewayS3TextractVercelDockerGitHub ActionsPrometheusn8n