AI Backend Engineer

Who Am I • Open Source • Competitive Programming • Projects • Tech Stack • GitHub

Who Am I?

AI Backend Engineer building and operating production systems handling 1M+ events/day and 50M+ DAU. I ship high-throughput services in Go and Python, optimize queries over millions of rows, and design APIs that stay fast under load.

Focused on measurable impact: ~40% query latency reductions and sub-50ms p50 API gateways.
Active open-source contributor to DeepSpeed, Cal.com, and LangGraph, fixing real production issues in widely used systems.

Current focus

• Distributed systems and event pipelines
• Multi-agent AI for real products
• Performance tuning across DB, cache, and network layers

---

Open Source Contributions (Selected)

DeepSpeed (Microsoft Research)

High-performance training infrastructure used at scale for large models.

• Fixed deadlocks in DecoupledCheckpointEngine during checkpointing failures
  Prevented indefinite training hangs by adding timeouts, process health checks, and robust cleanup logic, replacing unsafe assertions and ensuring checkpoint reliability even when subprocesses crash.
  PR #7742 · 🟣 Merged

• Prevented NaN propagation in OneBitLamb with empty parameter tensors
  Fixed a numerical edge case where sqrt(numel) caused 0.0/0.0 → NaN, corrupting the global scaling coefficient and destabilizing all optimizer state; guarded the denominator to ensure safe scaling.
  PR #7736 · 🟣 Merged

• Fixed runtime crash in NebulaCheckpointEngine during checkpoint commit
  Resolved an API mismatch by passing CheckpointCommitInfo instead of a raw tag string, unblocking Nebula-based checkpointing without breaking TorchCheckpointEngine.
  PR #7740 · 🟣 Merged

• Prevented runtime crash with PEFT / LoRA-wrapped Hugging Face models
  Fixed an incompatibility between DeepSpeed attention and PEFT adapters that caused crashes during training initialization.
  PR #7737 · 🟣 Merged

• Restored correct LR scaling under dynamic batching
  Fixed sqrt computation on non-tensor inputs to prevent TypeError and training instability when using dynamic batch sizes.
  PR #7735 · 🟣 Merged

---

Cal.com

Open-source scheduling infrastructure powering production teams and enterprises.

• Unblocked organization signup with pre-existing usernames
  Fixed a constraint in the org signup flow that prevented enterprise users from onboarding with reserved or existing handles.
  PR #25941 · 🟣 Merged

---

Featured Projects

NevraAI — AI-generated podcast pipeline
End-to-end system converting topics → scripts → voice → audio with async processing and vector search.
(FastAPI, Next.js, PostgreSQL, Redis, Qdrant, HuggingFace, GCP TTS)
Demo: Live · Repo: GitHub

OpenSkill (Open Source) — Claude skill manager CLI Command-line tool
CLI framework for defining reusable LLM skill modules with multi-provider support
(Go, Cobra CLI, LLMs (Groq, OpenAI, Anthropic, Ollama), Next.js)
Demo: Live · Repo: OpenSkill

SynthForce — Multi-agent startup simulator
AI agents collaborate to generate MVP roadmaps and market simulations in real time.
(FastAPI, Groq, MongoDB, WebSockets, Next.js)
Demo: Live · Repos: Backend | Frontend

VantageEdge — Low-latency API gateway
<10ms p50 gateway with rate limiting, load balancing, and real-time SSE analytics.
(Go, Redis, PostgreSQL, Next.js)
Demo: Live · Repos: Backend | Frontend

SyncLayer — Real-time collaborative board
Multi-user editing with WebSockets, Redis Pub/Sub, role-based access, and activity logs.
(Go, PostgreSQL, Redis, Next.js)
Demo: Live · Repos: Backend | Frontend

SentralQ — AI API debugger
Agentic system diagnosing auth, schema, and network failures with executable fixes.
(LangGraph, FastAPI, Next.js, Groq)
Demo: Live · Repos: Backend | Frontend

Aegis — Autonomous code review agents
LLM-powered agents analyzing codebases for architecture, security, and quality issues.
(Python, FastAPI, Redis, Groq Llama 3.3)
Demo: Live · Repos: PR Agent | Analyzer

Slanine — Production SaaS platform
25+ tools, Stripe payments, Redis caching, sub-200ms APIs.
(Next.js, PostgreSQL, Docker)
Demo: Live · Repo: GitHub

Qme — Community platform
Posts, voting, media uploads with Redis caching (-42% load time).
(Next.js, Redis, MongoDB)
Demo: Live · Repo: GitHub

Technical Arsenal

const rakshit = {
  backend: {
    languages: ["Go", "Python (FastAPI)", "Node.js (Express)"],
    patterns: ["Microservices", "Event-Driven", "CQRS", "Saga"],
    apis: ["REST", "GraphQL", "gRPC", "WebSockets"]
  },
  databases: {
    sql: ["PostgreSQL", "MySQL"],
    nosql: ["Redis", "MongoDB", "DynamoDB"],
    search: ["Elasticsearch/OpenSearch", "Qdrant"],
    optimization: ["Query Rewriting", "Materialized Views", "Connection Pooling"]
  },
  frontend: {
    frameworks: ["Next.js", "React", "Vue.js"],
    styling: ["Tailwind CSS", "Shadcn UI"],
    state: ["Zustand", "TanStack Query"]
  },
  cloud: {
    aws: ["EKS", "ECS", "Lambda", "S3", "API Gateway", "OpenSearch"],
    gcp: ["Cloud Run", "Compute Engine", "BigQuery", "Cloud TTS", "Pub/Sub"],
    other: ["Vercel", "Render"]
  },
  devops: {
    container: ["Docker", "Kubernetes"],
    ci_cd: ["GitHub Actions", "CircleCI"]
  },
  ai_ml: {
    frameworks: ["LangChain", "LangGraph", "LlamaIndex"],
    models: ["Groq", "GPT‑4o", "Llama"],
    rag: ["Qdrant", "Chroma", "Hybrid Search"],
    agents: ["Multi-Agent Systems", "Tool Calling", "ReAct"]
  }
};

---

GitHub Analytics

---

"It’s not who I am underneath, but what I do that defines me."

---