Prakhar Jain

AI Evaluation Engineer · Full-Stack Engineer · Co-Founder, Sarveio

India

Explore the 3D world

About me

AI evaluation engineer and full-stack developer who builds RL environments, Terminal-Bench-style tasks and grading rubrics to stress-test frontier LLMs (Claude Opus 5, Claude Fable 5.1), and ships production software: multi-tenant SaaS, real-time systems, 3D/WebGL experiences and cross-platform mobile apps. Co-founder of Sarveio (restaurant SaaS, $2,000+ revenue) and founder of Vernier, an AI benchmarking platform. Skilled in Python, TypeScript, React/Next.js, Node.js/FastAPI, Flutter and cloud delivery (Docker, Vercel, Cloudflare). Hackathon finalist and Algoverse Hackathon winner.

Experience

AI Data Annotator

Perit AI · Aug 2026 – Present

  • Drive model training data curation: decide which datasets and examples enter training and for which capability or task, directly shaping model quality.
  • Evaluate, rank and annotate LLM responses against detailed guidelines for accuracy, instruction-following, reasoning quality and safety.
  • Write prompts and reference answers and run side-by-side model comparisons (preference ranking) that feed training and evaluation.
  • Run quality checks and error analysis on data batches before training to keep labelling consistent and data clean.
  • LLM evaluation
  • Data annotation
  • Preference ranking
  • Prompt engineering
  • Error analysis

AI Evaluation Engineer

Cognyzer · Feb 2026 – Jul 2026

  • Engineered reinforcement learning (RL) environments and Terminal-Bench-style tasks to stress-test frontier LLMs and agents, with deterministic setup, programmatic verifiers and reproducible scoring.
  • Delivered 76 accepted, review-approved tasks in a single month, authoring multi-step command-line and agentic tasks that met strict quality and difficulty standards.
  • Benchmarked tasks against frontier models, primarily Claude Opus 5 and Claude Fable 5.1; analysed failure modes and wrote structured feedback to calibrate difficulty and surface capability gaps.
  • Designed grading rubrics and acceptance criteria that made model scoring consistent and objective across reviewers.
  • Leveraged Snorkel-based workflows to create, label and validate datasets against quality guidelines.
  • RL environments
  • Terminal-Bench
  • Rubric design
  • Snorkel
  • Python
  • Docker

Co-Founder & Full-Stack Engineer

Sarveio · Jan 2026 – Present

  • Co-founded and engineered Sarveio, a multi-tenant restaurant operating system (QR guest ordering, POS, KDS, inventory, CRM & analytics) on a MERN-based stack (React, TanStack Start, Express, MongoDB Atlas), taking it from architecture to $2,000+ in generated revenue.
  • Designed a multi-tenant data layer across 23 MongoDB collections with strict tenant isolation, atomic order-counter increments, and idempotency-key-protected order placement to eliminate duplicate charges under concurrent load.
  • Engineered a layered CDN and caching strategy (Cloudflare edge caching, ETag revalidation, stale-while-revalidate headers and in-memory request coalescing) serving public menu pages with fast, cookie-less, cacheable responses.
  • Built real-time order and kitchen-display sync with Pusher Channels and Socket.IO, backed by JWT authentication (httpOnly cookies), 5-role RBAC, rate limiting and CSRF protection to secure a production multi-tenant SaaS.
  • Automated WhatsApp order confirmations and marketing messaging via a Puppeteer-driven browser automation pipeline, and shipped WebXR/AR menu-item previews using Three.js and ARCore/ARKit.
  • React
  • TanStack Start
  • Express
  • MongoDB Atlas
  • Cloudflare
  • Pusher
  • Socket.IO
  • Puppeteer
  • Three.js
  • WebXR

Projects

Vernier

Founder · 2026

Benchmark-as-a-Service: on-demand benchmark environments, tasks and evaluation infrastructure for frontier models and AI agents.

  • Founded Vernier, a Benchmark-as-a-Service (BaaS) platform, applying hands-on RL-environment, rubric and task-validation experience to the cost and fragmentation of building reproducible benchmarks.
  • Shipped the live product site and an interactive evaluation dashboard UI with Next.js 16, React 19 and Tailwind v4, including scroll-linked Framer Motion walkthroughs, an accessible focus-trapped dialog, reduced-motion support and GitHub-connected Vercel deploys.
  • Architected the backend: a multi-tenant PostgreSQL schema (orgs, API keys, immutable benchmark revisions, runs, trials, artifacts, usage events), idempotent run creation, Redis rate limiting and S3/R2 trace storage with signed URLs.
  • Designed durable evaluation workflows (admit, boot, execute, collect, grade, finalize) with fair-share scheduling, per-plan concurrency caps, isolated sandbox workers (browser and shell pools) and deterministic, replayable graders.
  • Defined the public API and Python SDK contract (Benchmark(...).evaluate(agent, parallel=256)), plus Stripe usage metering, OpenTelemetry observability and a CI plan with Playwright.
  • Next.js 16
  • React 19
  • TypeScript
  • Tailwind CSS v4
  • Framer Motion
  • Vercel
  • PostgreSQL
  • Redis
  • S3/R2
  • Python SDK

Sarveio

Co-Founder & Full-Stack Engineer · 2026

Multi-tenant restaurant operating system: QR guest ordering, POS, kitchen display, inventory, CRM and analytics.

  • Took it from architecture to $2,000+ in generated revenue.
  • Multi-tenant data layer across 23 MongoDB collections with strict tenant isolation and idempotency-key-protected order placement.
  • Layered CDN and caching (Cloudflare edge, ETag revalidation, stale-while-revalidate, request coalescing) for fast, cacheable public menu pages.
  • Real-time order and kitchen-display sync with Pusher Channels and Socket.IO, secured with JWT, 5-role RBAC, rate limiting and CSRF protection.
  • WhatsApp order confirmations via Puppeteer automation, and WebXR/AR menu-item previews with Three.js.
  • React
  • TanStack Start
  • Express
  • MongoDB Atlas
  • Cloudflare
  • Pusher
  • Socket.IO
  • Three.js
  • WebXR

Portfolio World

Solo developer · 2026

This site: a playable Minecraft-style 3D voxel portfolio game with a custom engine.

  • Custom engine (fixed-step physics, camera rig, collision) and a Node.js pipeline converting Minecraft Anvil/NBT map data into a compact binary voxel format, meshed in Web Workers with greedy meshing and baked ambient occlusion.
  • Held 120 fps on desktop within hard budgets (max 250 draw calls / 2M triangles) using culling, quality tiers, half-float geometry (about 34% less GPU memory) and zero-allocation render loops.
  • Custom GLSL shaders and a post-processing stack (SSAO, bloom, god rays, depth of field); all audio generated procedurally with the Web Audio API.
  • Accessibility (keyboard-only, reduced motion, touch controls, pre-rendered no-WebGL fallback) and quality via 97 unit tests, 7 Playwright E2E tests and CI.
  • TypeScript
  • Three.js
  • WebGL2
  • GLSL
  • Web Workers
  • Web Audio API
  • Vite
  • Vitest
  • Playwright
  • GitHub Actions

Cura

Developer

AI health companion app for medical report interpretation, symptom Q&A and everyday health guidance.

  • Cross-platform health app with AI-powered medical report interpretation, symptom Q&A, diet and exercise guidance, appointment management and family connect.
  • Clean mobile UI/UX and AI assistant flows backed by a secured API proxy to keep keys off the client.
  • Flutter
  • Dart
  • Firebase
  • Riverpod
  • Claude API

Skills

AI Evaluation & LLMs

  • LLM evaluation
  • RL environments
  • Benchmark and task design
  • Rubric design
  • Terminal-Bench
  • Snorkel
  • Data annotation and QA
  • Preference ranking
  • Prompt engineering
  • Anthropic Claude API
  • Multi-provider LLM routing (OpenRouter, Gemini, Groq, Cerebras, Hugging Face)
  • Vector embeddings
  • OpenCV
  • Playwright browser automation
  • Webhook & cron-based automation

Languages

  • Python
  • TypeScript
  • JavaScript
  • Dart
  • SQL
  • GLSL
  • HTML5
  • CSS3

Frontend

  • React 18/19
  • Next.js (App Router, SSR/SSG, RSC)
  • TanStack (Start, Router, Query)
  • Vite
  • Redux / Zustand
  • Tailwind CSS
  • Radix UI / shadcn/ui
  • react-hook-form
  • Zod
  • Recharts
  • Framer Motion
  • GSAP
  • Three.js / React Three Fiber
  • WebGL2
  • WebXR

Backend & APIs

  • Node.js
  • Express.js
  • FastAPI
  • REST APIs
  • WebSockets (Socket.IO, Pusher)
  • Server-Sent Events
  • Celery
  • MERN Stack

Mobile

  • Flutter
  • Dart
  • Riverpod
  • Capacitor
  • Firebase (Auth, Firestore, Storage, Cloud Functions, FCM)

System Design

  • Multi-tenant architecture
  • Idempotent APIs
  • Durable workflows and queues
  • Fair-share scheduling
  • Sandboxed code execution
  • Object storage and signed URLs
  • Usage-based billing
  • Event-driven design
  • Caching and CDNs

Databases

  • MongoDB (Atlas, Mongoose ODM)
  • PostgreSQL
  • pgvector
  • Firebase Firestore
  • Redis

Cloud, DevOps & Testing

  • Docker & Docker Compose
  • Vercel
  • Cloudflare (CDN, Workers, Pages)
  • AWS S3 / R2
  • GitHub Actions (CI/CD)
  • HTTP/CDN caching (ETag, stale-while-revalidate)
  • Vitest
  • Playwright
  • Cloudinary

Web Platform & Performance

  • Web Workers
  • Web Audio API
  • GLSL shaders
  • Profiling
  • Frame and memory budgets
  • Accessibility (ARIA, reduced motion)

Geospatial & Maps

  • Leaflet
  • OpenStreetMap
  • Overpass API
  • Haversine distance and bearing calculations

Auth & Security

  • JWT
  • OAuth 2.0
  • Clerk
  • RBAC
  • bcrypt
  • Helmet
  • CSRF protection
  • Rate limiting

Tools & Design

  • Git / GitHub
  • Figma
  • Postman
  • ESLint / Prettier
  • VS Code
  • Cursor

Education

Bachelor of Engineering (B.E.), Computer Science & Engineering (Blockchain)

S.A.T.I (Samrat Ashok Technological Institute), Vidisha · Expected July 2028

CGPA: 7.6 / 10

Achievements

Algoverse Hackathon winner

Algoverse

Participated in 5+ hackathons, reaching the finalist round in several and winning the Algoverse Hackathon.

Co-founded and scaled Sarveio

Sarveio · 2026

Took a multi-tenant restaurant SaaS platform from architecture to $2,000+ in generated revenue.

6+ production apps shipped

Independently designed, built and shipped 6+ production web and mobile applications spanning e-commerce, AI/automation, 3D/WebGL and health-tech, across full-stack, mobile and AI/ML engineering.

76 accepted evaluation tasks in one month

Cognyzer · 2026

Authored multi-step command-line and agentic tasks that met strict quality and difficulty standards for frontier-model evaluation.

Contact

Email or LinkedIn is the fastest way to reach me.