Applied AI · Backend systems · Full-stack engineering
Computer Engineering student at NMIMS, Mumbai.
I work on retrieval-augmented generation (RAG), AI agents, and the APIs and interfaces around them. The part I enjoy most comes after the demo looks convincing: finding out where it breaks and what the evidence actually says.
Email me · LinkedIn · Selected work
For backend and product engineering, start with CARRICK and UdyogSaarthi. For AI/ML work, start with Graph-RAG. Follow the links to inspect the implementation and evidence.
CARRICK — Field reports into reviewed schedule changes
A runnable Python application that connects infrastructure site updates to schedule activities. AI proposes matches; planners review the actual dates and activities before export.
- Engineering: project-level access controls, retained source evidence, conflict checks, and validated XER/CSV export workflows. XER output passes a parser round trip; independent Oracle P6 import remains an external acceptance check.
- Stack: Python, SQLite, optional RAG with embeddings and structured extraction, browser-based capture and review.
System architecture · AI retrieval design · Verification suites · API implementation
UdyogSaarthi — Software for rural entrepreneurs
A team project: a multilingual platform that guides a business idea through feasibility, finance, compliance, and a Detailed Project Report.
- Engineering: server-side financial calculations, JWT-based applicant/reviewer/auditor roles, geospatial feasibility, and asynchronous PDF generation.
- Stack: Python, FastAPI, PostgreSQL/PostGIS, Redis, Celery, React, TypeScript.
System design · API contract · Backend tests · CI workflow
Graph-RAG — Can a retrieval system earn its confidence?
Co-author of a research project comparing multimodal, graph-based retrieval with a chunk-based RAG baseline.
- Engineering: PDF/text ingestion, knowledge-graph construction, grounded generation, and an evaluation pipeline covering answer quality, hallucination metrics, latency, and significance tests.
- Stack: Python, Neo4j, ChromaDB, LLM APIs.
Architecture · Evaluation methodology · Experiments
A result worth looking at — including the trade-off
The repository reports these aggregate benchmark results:
| Metric | GraphRAG | Chunk-based baseline |
|---|---|---|
| Hallucination rate | 0.33% | 2.04% |
| Mean response time | 6.00 s | 4.02 s |
| Semantic similarity | 0.5881 | 0.8308 |
Lower hallucination rate came with higher latency and lower semantic similarity in this experiment. Results depend on the corpus, metric definitions, and evaluation setup. Retrieval F1 uses different definitions across the pipelines and should not be treated as a direct comparison.
Also: AtlasAI — AI/ML lead on a team logistics prototype, combining a LangGraph agent, outcome-aware ChromaDB memory, human approval, and a NumPy LSTM for carrier forecasting. Its fleet dashboard uses simulated vessel positions.
The First Six Overs uses R and Cricsheet ball-by-ball data to analyse 1,227 completed IPL matches. Hypothesis tests and logistic regression examine toss effects, powerplay runs, and wickets. Cricket opinions are plentiful; I wanted a reproducible answer.
Read the findings · Run the analysis
If you're hiring for applied AI, backend, or full-stack software engineering, I'd like to hear about the problems your team is working on.
[email protected] · LinkedIn · All repositories
Outside the code: music, currently Daniel Caesar. Some cricket questions apparently require an entire repository.