Skip to content
Software Engineer at Bug0Silvassa, India

Saurav Jha

I build AI agents and RAG pipelines, and the harness that keeps them honest.
Tool calling, memory, guardrails, evals, and a cost line on every request.

Software Engineer at Bug0, where the agents I build write, run and repair end-to-end test suites on their own. Before that, 6 production apps and the real-time systems under them. I write up whatever I take apart along the way.

Read the Case StudiesResume
Saurav Jha
2.4
years shipping to production
4
AI systems in production
9
live products you can open
20
engineering posts

Five systems, and what was actually hard about each

Every card opens into the problem, the decisions with their reasons attached, the numbers, and what it still cannot do. The limits are there on purpose, because a portfolio without them is a brochure.

pragatiLM

live

A research notebook that answers only from your sources, with a locator on every claim

A single vector lookup on the raw question fails in two places that look identical from the outside. The question can be a bad search key, and whatever comes back is normally accepted without anyone asking whether it was enough to answer with.

LiveSource
5
query variants per turn
k=60
RRF constant
6 / 3
retry / refuse floor
900 / 150
chunk / overlap tokens
  • TypeScript
  • Express
  • Next.js
  • LangChain
  • OpenAI
  • Qdrant
  • PostgreSQL
  • Redis
  • Docker
  • Better Auth

Yugati

live

An agent with real Gmail and Calendar tools, plus the admin panel that watches what it spends

Handing an LLM write access to somebody's inbox is the easy part. Knowing what it did, what it cost, and whether the last message was an attempt to talk it into something is the part that decides whether the thing can be charged for.

LiveSourceWalkthroughDocs
51
tests in CI
40
turn session cap
2
guardrails, input and output
INR
live paid billing
  • Next.js
  • TypeScript
  • OpenAI Agents SDK
  • tRPC
  • Drizzle
  • PostgreSQL
  • Redis
  • TanStack Query
  • Tailwind CSS
  • Vercel

FormCraft

live

A form builder with real undo, autosave and response analytics, typed end to end

The field list is the easy half of a form builder. The hard half is keeping the canvas responsive while every keystroke is persisted, reversible, and type safe from the database through to the browser.

LiveSource
14
field types
50
step undo history
1.2s
autosave debounce
4
stage form lifecycle
  • Next.js
  • TypeScript
  • tRPC
  • Zustand
  • dnd-kit
  • Clerk
  • PostgreSQL
  • Recharts
  • Tailwind CSS

Voxly

live

Real time polling where one link is the form while it is open and the results once it closes

A poll has two audiences that want opposite things. The creator wants tallies moving as answers land, the respondent wants a page that does not shift under them. Both of them arrive at the same URL.

LiveSource
1
link for form and results
3
step creation wizard
410
returned on a late submit
2
response modes, anon and signed in
  • React
  • Node.js
  • Express
  • PostgreSQL
  • Socket.io
  • Clerk
  • Recharts

CodeWarriors

live

A DSA practice platform that runs untrusted submissions against hidden tests and tracks the progress

Executing code written by strangers is the part that decides whether a coding platform is a demo or a product. Almost everything else around it is ordinary CRUD.

LiveSource
9
feature areas, dashboard to admin
2
paths: run on samples, submit on hidden
5
GitHub stars, most starred repo
  • React
  • TypeScript
  • Vite
  • Redux Toolkit
  • Monaco
  • Node.js
  • Express
  • Prisma
  • PostgreSQL
  • Judge0
  • Cloudflare Workers
  • OpenAI

Everything else that is public

All repositories

Four more repositories worth opening. The full list runs to 119.

  • NoteCast

    AI

    The straightforward hybrid of the retrieval work: PDFs, articles and YouTube into one Qdrant index, with every answer anchored to the chunk it came from.

    • Next.js
    • Vercel AI SDK
    • LangChain
    • OpenAI
    • Qdrant
    • Prisma

  • ThumbAI

    AI

    Thumbnails from a sentence. Prompts are rewritten and guardrailed first, then jobs run behind a webhook queue that streams previews over SSE.

    • Next.js
    • OpenAI
    • Fal AI
    • Google Gemini
    • PostgreSQL
    • Prisma

  • custom-oidc

    An OIDC provider written from the spec rather than configured, which is where the blog post came from.

    • TypeScript
    • OAuth 2.0
    • OIDC

  • UIgnite

    A customisable React component library published as a Lerna monorepo.

    • React
    • Radix UI
    • Framer Motion
    • Lerna

Skills, with the work that backs them

42 of these 54 name the project they were used in. The rest are things worked with but not shipped in anything public, and they say so by staying blank.

Retrieval

Getting the right context in front of the model, and knowing when it is not there.

  • RAGpragatiLM, NoteCast
  • HyDEpragatiLM
  • Query rewritingpragatiLM, ThumbAI
  • Step back promptingpragatiLM
  • Reciprocal rank fusionpragatiLM
  • RerankingpragatiLM
  • Hybrid searchpragatiLM
  • Semantic chunkingpragatiLM, NoteCast
  • CRAG self correctionpragatiLM
  • Retrieval eval, hit rate and MRRpragatiLM

Agents and guardrails

Letting a model take actions, with something watching what it does.

  • OpenAI Agents SDKYugati
  • Tool and function callingYugati, Bug0
  • MCP
  • Prompt injection classificationYugati
  • PII output filteringYugati
  • Structured outputspragatiLM
  • Conversational memoryYugati, pragatiLM
  • Voice agentsYugati
  • Token and cost accountingYugati
  • Self healing test agentsBug0

Data and vectors

Where the embeddings, the rows and the hot state live.

  • QdrantpragatiLM, NoteCast
  • Pinecone
  • PostgreSQLmost projects
  • MongoDBGamazing
  • MySQLGamazing
  • RedispragatiLM, Yugati
  • Valkey
  • PrismaNoteCast
  • DrizzleYugati
  • Postgres full text searchpragatiLM

Backend and runtime

The services around the model, and the streams that carry its output.

  • TypeScripteverything
  • Node.js
  • Python
  • ExpresspragatiLM
  • Fastify
  • tRPCYugati
  • Server sent eventsYugati, pragatiLM, ThumbAI
  • WebSocketsVoxly, Gamazing
  • Queue workerspragatiLM, ThumbAI
  • C++

Interface

Streaming tokens, citation viewers and admin panels are frontend problems too.

  • Next.js App Routerall recent work
  • React 19
  • Tailwind CSS
  • TanStack QueryYugati
  • Framer MotionUIgnite, this site
  • Three.jsthis site
  • RechartsYugati

Delivery

Whether it runs on someone else's machine, and whether a regression is caught.

  • DockerpragatiLM
  • GitHub ActionsYugati
  • PlaywrightBug0
  • VitestYugati
  • Turborepo
  • Vercel
  • Git

Where the practice came from

Two roles. The first taught real-time systems under load, the second is where the agent work started.

  1. Nov 2025
    to Present

    Hybrid, India

    current

    Bug0

    Software Engineer

    AI driven QA automation.

    • Builds AI agents that autonomously generate, execute and repair end to end Playwright suites through LLM tool calling.
    • Self healing tests that detect DOM and selector drift and regenerate the script at runtime rather than failing the build on a moved button.
    • Wired into CI/CD so regression coverage runs on every build.
    • TypeScript
    • Playwright
    • LLM tool calling
    • CI/CD

  2. Apr 2024
    to Oct 2025

    Remote

    Gamazing Play Studios

    Full Stack Developer

    Six production apps and the real-time systems underneath them.

    • Shipped 6+ production apps with admin dashboards: grocery, food delivery, rider, cafe and real money gaming.
    • Node.js and Express backends on MongoDB and MySQL, with secure REST and WebSocket APIs.
    • Low latency real-time systems over Socket.io for live gameplay, crash and bidding rounds, order tracking and push notifications.
    • Puppeteer scrapers ingesting third party inventory.
    • Node.js
    • Express
    • MongoDB
    • MySQL
    • Socket.io
    • Puppeteer

  3. 2019
    to 2023

    Dr. Jivraj Mehta Institute of Technology

    B.Tech, Computer Engineering. CGPA 9.06 / 10

Writing, mostly about the layer underneath

All posts

Twenty posts on Hashnode. The pattern is the same as the work: take something people use without looking inside, then open it.

Open to Applied AI Engineer roles

If you are building agents or retrieval systems and want someone who will argue about thresholds and then write the eval harness, get in touch. Fastest by email.

Send an EmailResume

Elsewhere