Building AI systems, backend infrastructure, and reliable software for real-world problems.
I'm a Computer Science undergraduate focused on Software Engineering, Applied AI, backend systems, and reliability.
I’m especially interested in projects where the hard part is not just getting something to work once, but making it correct, inspectable, failure-aware, and useful to real users.
My work spans:
- AI systems with deterministic safety boundaries
- backend and reliability engineering
- developer tooling
- applied machine learning
- research
- data structures and algorithms
| Project | What it is | Stack | Highlight |
|---|---|---|---|
| Amana | Agentic commerce gateway that turns existing merchant systems into AI-transactable businesses and enables a safe AI buyer | Java · Spring Boot · Next.js · PostgreSQL · pgvector · Gemini · Razorpay | 250/250 deterministic safety cases passed · fail-closed transaction authority |
| RepoPilot | Local-first coding-agent platform built to understand repositories before changing them | Python · FastAPI · tree-sitter · BM25 · Ollama | Structure-aware repository retrieval · local inference · human-in-the-loop architecture |
| HookRelay | Reliable webhook delivery system built around explicit failure handling and recovery | TypeScript · Fastify · PostgreSQL · Redis · BullMQ | Idempotency · retries · HMAC signing · DLQ · replay · at-least-once delivery |
| VoteReady | Citizen-first household navigator for India's SIR voter process | Next.js · TypeScript · Gemini · Supabase · PostgreSQL | Multilingual record extraction · deterministic matching · human confirmation |
| PS-Prep | LeetCode-style programming-prep platform built for my college's Project School selection test | React · Vite · Supabase · Gemini · Monaco Editor | ~100+ student users · improved from real user feedback |
| Skin Lesion Classification | Error-driven HAM10000 skin-lesion classification study | PyTorch · TensorFlow · ResNet50 · Grad-CAM | 57.29% → 84.76% test accuracy with class-imbalance-aware training |
AI handles meaning. Deterministic software controls truth, authority, and money.
Amana separates AI reasoning from financial authority.
The AI can understand intent, inspect merchant APIs, diagnose failures, and suggest mappings, while deterministic software owns the transaction path:
proposal → authorization → execution → payment truth
The system includes Razorpay Test Mode payments, webhook verification, reconciliation, bounded execution, fail-closed checks, and a deterministic adversarial safety proof.
RepoPilot is being built around the idea that a coding agent should understand repository structure before attempting a patch.
Currently implemented:
- repository discovery and ingestion
- Python structure-aware parsing
- semantic code chunks with source provenance
- lexical code retrieval
- local Ollama provider boundary
- automated parser, retrieval, pipeline, and API tests
Target direction:
Repository + Issue
↓
Understand Structure
↓
Retrieve Evidence
↓
Plan
↓
Human Approval
↓
Patch
↓
Test
↓
Critique
The goal is an inspectable engineering agent, not an unrestricted autonomous coding bot.
PS-Prep started from a problem my classmates and I were actually facing.
Preparation material for our Project School selection test was scattered across old resources, coding environments, solutions, and chat tools, so I built a single platform for practice.
I shared it across multiple college classes during exam season, where it was used by roughly 100+ students.
One of the first user-reported problems was that programs using input() had no convenient way to receive values. I added a dedicated stdin workflow based directly on that feedback.
Build → ship → get feedback → fix the real problem.
I'm involved in applied research across two domains:
Research work through IIIT-H / iHub-Data, involving computer-vision experimentation and evaluation.
Research with Drugparadigm on bispecific-antibody manufacturability, involving large biological sequence datasets and constrained evaluation settings.
I'm particularly interested in ML problems involving:
messy data · limited labels · domain constraints · evaluation uncertainty
|
Models are useful for ambiguity, language, diagnosis, and planning. Deterministic software should own irreversible authority and correctness-critical decisions. |
Retries, idempotency, reconciliation, concurrency, and recovery belong in the system design from the beginning. |
|
I prefer measured behavior, failure analysis, tests, and reproducible evidence over a system that only worked once. |
The projects I value most usually start with somebody genuinely needing the thing. |
- building deeper backend, AI-systems, and reliability projects
- working on applied ML research
- practicing DSA and competitive programming
- preparing for Software Engineering / Applied AI internships
I'm particularly interested in opportunities involving AI systems, backend infrastructure, reliability, developer tooling, and applied machine learning.