I am an AI engineer in Mumbai with 3+ years on production LLM systems. I run those systems in production and I build the whole product around them.
I came to the AI work from full stack work on the same products, so I still build the whole thing: Python, FastAPI, TypeScript, React, Next.js and PostgreSQL.
I am open to AI Engineer, Applied AI Engineer, GenAI Engineer, Forward Deployed Engineer and Software Engineer II roles.
Email: sk729584@gmail.com
Portfolio: saurav02022.github.io
LinkedIn: linkedin.com/in/saurav02022
- Build LLM systems that make a real decision, so a person checks the output instead of producing all of it.
- Measure what the model decides against human judgement, and gate every change on that number.
- Keep LLM cost and latency down with response caching and queued inference.
- Build the whole stack around it on FastAPI, PostgreSQL, Redis and Next.js, including CI/CD and deployment.
- Build offline-safe flows for schools where the internet is weak, so nothing is lost when the connection drops.
Shikha Learning Labs is an EdTech initiative of the Shantilal Shanghvi Foundation. We build digital products for schools, teachers, students, principals and school admins.
I work as an AI Software Engineer. Our products are used across 117 schools. I am the sole engineer for 4 of 10 web products and I also work on shared parts of the platform.
My work includes:
- Built a rubric-scored analyser that takes the first pass on 5,200+ student submissions a month, scored against per-class, per-subject rubrics. It returns approve, reject or a comment, so teachers cross-check instead of reading every one. First-pass teacher review work is down 64%.
- Measure that analyser against teacher judgement instead of assuming it is right. Agreement went from 52% to 81% on a 1,400-submission evaluation set.
- Set up an evaluation gate for LLM releases. Any prompt or model change is blocked when agreement falls below 78% or the teacher override rate crosses 18%, so regressions are caught before they reach teachers.
- Built the AI feedback workflow for classroom audio on FastAPI workers and Redis queues, with retries and idempotent writes. It carries 430+ recordings a week for 215 teachers and brings a mentor's review of one class from around 60 minutes to 5-10 minutes.
- Cut repeat LLM spend by 38% across 3,200+ cached results with a response cache keyed on a SHA-256 hash of the scores, the prompt version and the model, so a result is recomputed only when one of those changes.
- Migrated nine backend domains off the Supabase SDK onto async PostgreSQL with SQLAlchemy, asyncpg and Alembic, moving the database RPC logic into application code and cutting a course import from about 160 sequential calls to one atomic bulk write.
- Collapsed LMS course retrieval from 8 to 12 sequential queries into a single set-based read, writing characterization tests before the refactor and pinning the result with a query-count test that fails if a second query appears.
- Moved multi-tenant authorization into a Supabase custom access-token hook that resolves organization, role and workspace permissions into JWT claims. That removed a cross-region profile read from every request and closed organization-scoping gaps an access audit had found.
- Merged the separate frontend applications into one typed Next.js monorepo with shared authentication and a shared component library, generating the TypeScript API client from the backend's OpenAPI schema so a backend change that breaks the frontend fails CI.
- Brought the live GCP estate under Terraform with no downtime, writing import blocks for VPC, IAM, service accounts, Secret Manager, Redis, Artifact Registry and Cloud Scheduler, then built nine CI/CD pipelines so every service deploys the same way to development and production.
- Moved the team to token-based deploys from GitHub Actions across 12 projects, so six engineers ship from one paid Vercel seat and we save about $1,200 a year.
- Built offline-safe recording with IndexedDB for around 50 rural schools, so no class audio is lost when the connection drops.
- Built shared UI components, single login and role-based access used across products, so everyone gets the right access.
Two research papers from the Shikha team are connected to products I built or worked on.
These papers were written by Shikha's founder, product managers, researchers and wider team members. I am not listing myself as an author. My connection is through the engineering and product work behind these tools.
-
AI-Human Synergy: Using Design Thinking to Build for and with Teachers
Related to our AI teacher workflow and multi-assistant platform for teachers. -
Conversations for Learning: Designing Personified Historical Chatbots to Enhance Critical Thinking in K-12 Students
Related to AI chatbot work for K-12 learning and student questioning.
Nuveb is an open OTT network for creators. It helps creators publish video content and earn from it without depending only on large platforms.
I worked as a Full Stack Developer across the viewer-facing OTT platform and the creator side of the product.
My work included:
- Built the creator portal for this open OTT video network, used by 10,000+ creators, covering multi-step video upload, publishing, scheduling, earnings views and payment-status tracking.
- Took key viewer-app page load from 8.5 seconds to 2.5 for 50,000+ monthly users by profiling LCP and FCP and splitting rendering: personalised routes server-side, browse pages static.
- Cut interaction delay on image-heavy pages by 42% across a 10,000-item catalogue by reworking image loading, rendering behaviour and the frontend's use of Sharp-backed image APIs.
-
rto-shield
An ops console where an AI voice call confirms a cash-on-delivery order before it ships, resolving each completed call into one of six order states. The provider reports the same call up to three times, so one idempotent mutator keyed on the call ID decides the order state and a repeat can never ship a second parcel. 33 tests cover retries, repeat callbacks and state transitions, at 91% coverage of the core order flow. -
claims-processing-system
A health-insurance payout engine. The rules are about 170 lines of plain Python with nothing from the web framework in them, covered by 74 tests across 94% of rule-engine paths, and the whole claim write sits inside one PostgreSQL function. -
chat-ai-app A streaming chat app built on the Vercel AI SDK with OpenAI models. I have kept working on it since November 2024 — 208 commits across 201 separate days.
-
ai-interview
A mock interview platform with video calls, transcripts and AI feedback. -
resume-builder
A resume tailoring tool for one job, with changes shown side by side. Try it -
linkedin-hashtag-refresh-engine-app
A small tool that drafts hashtag sets for LinkedIn posts. Try it -
financial-literacy-assistant
A budgeting assistant for beginners, built in a three-hour hackathon.
- Mentored 41 contributors across the open-source projects EduFlow AI and VidyAI++ at Social Summer of Code 2026, and filed 5 tracked issues on VidyAI++ including 44 TypeScript errors the build was hiding.
- Contributed at Social Winter of Code 2026, with light and dark theming across 11 files in AlgoFi.
- Solved 200+ DSA problems on LeetCode and takeUforward. LeetCode contest rating: 1,616.
AI and LLM systems: LLM evaluation, prompt engineering, structured outputs, response caching, queued inference, AI voice agents, LLM APIs (OpenAI, Gemini, DeepSeek).
Languages: Python, TypeScript, JavaScript, SQL.
Backend: FastAPI, Node.js, REST APIs, OpenAPI, asynchronous processing, Redis, message queues, SQLAlchemy, asyncpg, Alembic, JWT.
Frontend: React, Next.js, Tailwind CSS.
Databases and infrastructure: PostgreSQL, Supabase, Firestore, Docker, Terraform, GitHub Actions, CI/CD, GCP, Google Cloud Run, Vercel.
Testing and observability: pytest, Vitest, Playwright, regression testing, Sentry, monitoring.
- MCA, Artificial Intelligence and Machine Learning, Indian Institute of Information Technology Ranchi, 2024-2026, GPA 8.0/10
- Full-Stack Web Development, Masai School, 2022-2023, GPA 9/10
- B.Sc. Mathematics (Honours), Munger University, 2019-2022