The interesting part is never the idea.

Every submission is untrusted.

An online judge executes code written by people you have never met. Each submission is C++ or Python, and any one of them can loop forever, eat all the memory, or simply fail to compile.

Hard limits, no exceptions.

Each submission runs in a sandbox with 1,000 ms of time and 256 MB of memory. Go past either and the judge answers TLE or MLE; fail to compile and it answers CE.

Around the engine: JWT auth with role-based access, a rate limiter on the submission API to shut down abuse, and a Gemini code reviewer that handles three review scenarios.

50+ active users, 500+ submissions.

The submissions came in during its first month. Judging and the web app ran as two services, on AWS EC2 with images in ECR and on Render, and verdicts came back in under 200 ms end to end. Three features cleared code review and shipped inside their sprints. The project is archived now, and the code is public.

The clock starts at checkout.

A 60-minute delivery window leaves no room for a slow API. Checkout, order tracking and routing all sit between a tap on a phone and a parcel at a door.

Separate services, cached reads.

I built the platform's core flows as RESTful microservices, each deployed to production independently: the Try-First-Buy-Later checkout, real-time order tracking, and the logistics routing pipeline.

I designed the MongoDB and PostgreSQL schemas behind them, deployed the services on AWS, and added Redis caching.

API latency down by more than 40%.

That came from the Redis layer. I also maintained the CI/CD pipelines and wrote the unit and integration test suites that every release went out behind.

MarketPulse-X

A supply chain risk platform, built at Google's Agentathon: a Guinness World Record event with 15,000+ participants. It was a team build. I owned the LangGraph orchestration and state management for the six agents; teammates built the SEC EDGAR integration and the GKE deployment.

When an analysis scores below 70% confidence, a loop re-runs the upstream agents instead of shipping the answer. The pipeline fuses supplier data parsed from SEC 10-K filings with live news, so risk is re-assessed as events arrive.

IISPPR

At the International Institute of SDGs & Public Policy Research I led a team of nine and built a three-agent pipeline in n8n for academic text. One agent rewrites. A second verifies the rewrite against the source. A third refines it with domain keywords while preserving the meaning.

I also extended the team's existing n8n automation workflows, working inside their architecture and conventions.

CloudSense

An AWS cost agent that learns from the recommendations you reject. It is built on Hindsight agent memory, so a "no" today changes what it suggests tomorrow. I wrote up what it taught me in an article, "Hindsight Taught My AWS Cost Agent to Leave Standbys Alone".

PlaceIQ

Placement management for the officers who run it at Tier-2 and Tier-3 colleges. It started as a job-alert app for students; I reframed it as infrastructure for coordinators, because that is where the work piles up.

Many colleges share one system, so access is a four-tier hierarchy across 65 REST endpoints, with optimistic concurrency control so two coordinators cannot overwrite each other. Students get a resume builder with a live ATS score and a LaTeX mode; coordinators get Kanban tracking.

Resume scoring keeps model cost bounded: rule-based matching scores every submission, and the LLM review runs only on demand. 22 Jest and Supertest tests run in CI.

Kairos

Interview prep for Indian engineering students, built like a habit instead of a syllabus: a short daily question feed, AI scoring with feedback, and a streak to protect.

I'm building it as a startup. Before adding features I'm running a four-week retention test, because the riskiest assumption is whether people come back tomorrow.

Arcanis

A competitive magic RPG that I am building as founder. You describe a spell in plain language, an LLM turns it into a balanced, classified spell with its own card art, and then players battle with the spells they wrote.

Spells are made in a Lab across a 12-element system. Battles are asynchronous PvP run by a rule-based engine, so a fight costs nothing in model calls. The hard part is fairness: anyone can type anything, and the game still has to stay balanced.

DevMem-Agents

My B.Tech minor project: an upgrade to the memory of Stanford's Generative Agents, the simulated town of Smallville. The flat memory stream is replaced by a four-stage developmental memory: personality priors, episodic memory, sleep-triggered consolidation, and identity. Everything else stays as upstream, so memory is the only variable.

I run it as project owner, with AI coding agents doing the implementation, under one hard constraint: zero API cost, using only free-tier keys rotated by a custom router. The baseline-versus-staged comparison is set up, and there are no results to report yet. The design is written up as an IEEE-format paper with two co-authors, presented in October 2026.

How memory should work

Underneath the projects sits a set of papers on memory in learning systems. The furthest along, S-DAM, asks whether seeding an associative memory with core-knowledge priors helps it retrieve. Its most useful result is a negative one: a feature can be decodable and still be useless to memory.