Conflict-Aware Graph RAG
Graph-based retrieval-augmented generation that turns text into knowledge triples with LLMs. An entropy-based conflict-detection module cuts hallucination on multi-hop question answering.
40 projects across AI/ML research, developer tools and the web, plus the open-source libraries I contribute to.
Graph-based retrieval-augmented generation that turns text into knowledge triples with LLMs. An entropy-based conflict-detection module cuts hallucination on multi-hop question answering.
India's first open Rajasthani-Hindi code-switched corpus of 50K sentences. Fine-tuned MuRIL beats GPT-4o on NER, sentiment, and toxicity detection.
140x compressed deep-learning model for real-time 5-class arrhythmia detection on an Arduino Nano 33 BLE. Hits ~99% F1 in under 256KB using spectral knowledge distillation and ECG+PPG sensor fusion.
A CNN encoder with an LSTM decoder that writes natural-language captions for images, trained on the Flickr8k dataset.
A fully AI-powered text + image RPG set in the Wizarding World. Combines an Ollama LLM narrator, NetworkX knowledge graph, emotion-driven storytelling, Q-learning enemy AI, and procedural scene rendering.
Multi-step prompt-injection attacks that push a tool-using LLM agent to exfiltrate data past its guardrails, built for Kaggle's AI Agent Security competition (best public score 106.285, with a public working note), plus a clean-room benchmark for guardrail bypasses and confused-deputy exfiltration.
Network intrusion detector combining Random Forest, XGBoost, and a neural network, with feature engineering and evaluation on standard IDS benchmark traffic.
A stand-in for the real LLM in your tests: record LLM API calls once, then replay them forever, fast, free, deterministic, and offline. Hooks httpx, requests, and aiohttp, with fuzzy and semantic matchers plus built-in secret redaction. Install with pip install standin.
A safety hook for AI coding agents: it logs every tool call and blocks file access outside the folders you allow, with a live dashboard. Works with Claude Code, Gemini CLI, and Cursor. Install with pip install sneakoscope.
Made the filesystem server's move_file fail instead of silently overwriting an existing destination file.
Three merged PRs: a failed read no longer leaves a poisoned in-progress entry in the client-side cache, CLIENT INFO no longer crashes on values containing '=' or spaces, and new unit tests for the response helpers.
The reader now recovers from a malformed PDF whose destination is not an array instead of crashing.
Fixed a crash in exp.cast() when the input is already a cast to INTERVAL.
A truncated or out-of-bounds cmap subtable header now raises a clear TTLibError.
QUAL files with an empty title line are handled without raising IndexError.
Escaped array-literal elements in the PostgREST contains, contained_by and overlaps filters so special characters build a valid query.
Out-of-range integers and pre-epoch timestamps now raise a clear error when encoding AMQP field tables, instead of a raw struct.error.
Inbound mail now handles envelope addresses that have no domain.
Coercing a non-finite decimal to bigint now raises a clear ValueError.
Fixed the token-bucket limiter in both the JavaScript and Python SDKs to reject requests that ask for more tokens than remain.
Two feature-flag matching fixes: exact and is_not compare case-insensitively after string coercion, and not_regex handles every value type the way regex does.
Plus fixes in trimesh, mutagen, RDFLib, tomlkit, fastavro, Typebot, WorkOS and many more, mostly crash-hardening of file-format and protocol parsers found through differential testing.
Kaggle Competitions Expert working across ML domains, with a silver medal in Kaggriculture (473rd of 10,148) and a bronze in Biohub cell tracking (287th of 3,947), plus medical imaging, coding agents and agent security. Read the full breakdown of how I approach them and what I have worked on.
Clean, minimal Python implementations of modern ML and AI research papers, collected in one repository.
PyTorch implementation of Mamba selective state space models for long-sequence modelling, with the selective scan and hardware-aware recurrence.
PyTorch implementation of RoFormer's rotary position embeddings for attention, with tests against the reference formulation.
Selective timestep weighting and advantage replay for more sample-efficient RLHF fine-tuning of diffusion models.
Neural Cellular Automata for architecture-agnostic weight generation, exploring self-organising updates as a learned optimiser.
A five-stage optimizer meta-pipeline unifying SGD, Adam, and Lion under one schedule for comparison and ablation.
Scalable linear-RNN state via sparse key-value deltas, trading dense recurrence for a compressed, updatable memory.
Open-source, local-first real-time meeting and interview copilot: on-device transcription with a live assist panel, an alternative to cloud tools like Cluely.
A hardened, zero-setup Model Context Protocol server for WhatsApp, letting agents read and send messages through a safe, well-scoped interface.
Finds source files that have no test and ranks them by risk. Static, zero-config, language-agnostic triage for where to add coverage first.
Feed it a GitHub profile and it roasts your commit history and repo choices. A small, shareable browser app.
Generates an architecture diagram from a repository, turning its structure and imports into a readable map.
A deliberately bad GPU memory allocator simulator: it allocates in the worst reasonable way so you can watch fragmentation and OOM behaviour up close.
An outbound AI voice agent that calls leads on its own, pitches in Telugu, Hindi or English, classifies them hot, warm or cold, sends a WhatsApp message mid-call when intent is high, and books callbacks from spoken times.
A private screen memory for macOS: records screen, input, audio and window context on-device, blocks sensitive apps at capture time, and makes it all searchable. Ships as a native app and a CLI daemon.
Squad optimizer for IPL cricket auctions using XGBoost player valuation and Integer Linear Programming for budget-constrained team selection on Cricsheet ball-by-ball data.
Satellite-based Urban Heat Island analysis of Ahmedabad (2013-2024) using Landsat 8, Sentinel-2, Google Earth Engine, and scikit-learn for LST mapping and hotspot prediction.
Mutual Fund Intelligence Dashboard for tracking fund manager holdings, portfolio overlap, style drift, SIP XIRR, and BHB performance attribution.
A from-scratch SQL and DBMS course for placement prep: runnable MySQL queries against a small bookstore database, plus the DBMS theory interviews test.
Reproductions of recent reversible-data-hiding papers (IEEE TMM, SPL, and TCE, 2024 to 2026): encryption, embedding, and extraction pipelines benchmarked for capacity and reconstruction fidelity.
Full e-commerce platform with a product catalog, shopping cart, checkout, user authentication, and order management.
The same store rebuilt on a PHP backend with MongoDB, keeping a React-compatible frontend.
A distributed file-storage system with node-based storage.
A full-stack web application built on the SERN stack.
A faithful replica of Charles Leclerc's website: a smooth-scroll, animation-driven experience.
A full-stack intelligence layer for the AI economy: a Next.js app tying graph data, APIs, and an interface into one workspace.
A metadata-driven app runtime: a JSON config compiles into a live UI, APIs, a database, and auth, so a whole app is described rather than hand-coded.
A full-stack Swiggy-style food delivery app on the MERN stack, with restaurant listings, cart, and order flow.
A clean weather app that pulls live forecasts from a public API, with city search and saved locations.
A CRUD system for managing books, members, and loans, with search, due-date tracking, and an admin dashboard.