Skip to content
Arjit Kulkarni
03Capabilities

What I build, and how I know it works.

I build across the stack: responsive React interfaces, dependable services, and AI features with evidence behind their decisions. Here is the toolkit I bring to SDE and AI engineering work.
01

Frontend Engineering

Frontends built with the same care as the services behind them — typed, accessible, and fast on the devices people actually use.

  • React.js
  • Next.js
  • TypeScript
  • Tailwind CSS
  • WebGL / three.js
02

Backend Services

REST APIs designed before they are written — clear resources, honest error states, and a data model that still makes sense six months in.

  • FastAPI, Node.js, NestJS
  • REST API design
  • Authentication
  • PostgreSQL schema design
  • Query optimisation
03

AI & Machine Learning

AI agents with typed tools and verifiable actions. Retrieval systems and reproducible ML experiments that turn a model into useful software.

  • AI agents & MCP tools
  • PyTorch
  • Transfer learning
  • Experiment design & tracking
  • Evaluation pipelines
  • Vector search & retrieval
04

Systems Programming

Work close to the machine: storage engines, concurrency, and the memory and I/O behaviour that decides whether a system holds under load.

  • C++17, RAII, ownership
  • Concurrency & multithreading
  • Caching, indexing, persistence
  • File I/O and durability
  • Storage-engine internals
05

Performance & Debugging

Measure, then change. Profiles and query plans decide what gets optimised — never a hunch about where the time is going.

  • Linux perf, profiling
  • PostgreSQL query plans
  • Benchmarking
  • Root-cause analysis
  • ThreadSanitizer, fault injection
06

Delivery & Tooling

Small commits, real code review, green CI. The unglamorous parts are what make shipping to production routine instead of eventful.

  • Git, GitHub Actions, CI/CD
  • Docker, CMake
  • Unit & integration testing
  • Prometheus, Grafana
  • Vercel, Neon PostgreSQL
07Process
The same four steps, from an interactive frontend to an AI workflow.
  1. 01

    Understand the user

    Start with the task someone needs to complete. Map the interaction, the data and the failure states so the interface and its services solve the same problem.

  2. 02

    Design the data first

    Schema and retrieval path together, not search bolted onto an existing model. A data model that survives six months of change is worth more than any amount of clever code above it.

  3. 03

    Ship it small

    Incremental delivery through Git, real code review and CI. Small changes are easier to review, easier to revert, and much easier to attribute when something moves in the wrong direction.

  4. 04

    Prove it held

    Profiles, query plans, benchmarks, load tests, fault injection and sanitizers. If the improvement cannot be measured after the fact, it was a guess that happened to feel faster.

See how this plays out in practice.

Case studiesStart a conversation