A privacy-safe collective operational memory for software vendor fleets — sanitized incident fingerprints feed a shared "HIVE" so a fix found at one deployment reaches every other. Finalist, Elastic × AWS Forge the Future 2026.

On-device ASL input accelerator on an iQOO 15 — sign a couple of glosses, get candidate sentences, and the phone speaks them into a call. MediaPipe geometry for recognition, a local LLM only for unseen sequences.

School-consolidation investigation tool for district education officers — compare receiving schools side by side, run eight schema-checked agents per candidate, and choose last. The LLM never produces numbers or approvals.

AI-organized before/after field evidence and traceable impact reports for NGOs, built on Cloudinary — semantic search, human-confirmed pairs and a full provenance trail for every image.

Matches India’s monthly CDSCO failed-batch alerts against hospital dispensing — including hand-typed and missing batch numbers — and lists exposed patients for clinical follow-up. Data layer built; Snowflake pipeline designed.

Describe an app, import your files, get an installable offline PWA. A service worker serves LLM-generated apps from IndexedDB so Chrome can install them — no backend of its own.

Built at a Sarvam AI hackathon, hosted on GrowthX Built. [details pending — description, tech stack, and your role to be confirmed]

Drop in a scientific paper. Get out an audio podcast summary, a slide deck, and a plain-language explainer — built to bridge dense academic literature and the people who need it.

Agentic pipeline using the Gemini API to generate, optimise, and orchestrate multi-step SQL queries — LLM-driven database reasoning with a modular Python architecture.

LLM-maintained knowledge base for research papers — ingests arXiv PDFs into a continuously-updated markdown wiki, tracing every generated answer back to the exact source passage instead of relying on typical RAG retrieval.

Bulk-generates formatted certificates from Excel data and a Word template. Built for college sports day, later found useful by strangers.

ESP32-based solar panel that dynamically adjusts tilt using computed sun position from latitude/longitude — no LDR sensor, pure math.

An MCP server deployed across multiple European hospitals, exposing 15+ AI tools through FastAPI with LangGraph orchestration, plus an LLM analytics layer with a Deep Research mode that has served 15,000+ queries. RAG-based tool selection over ChromaDB with model routing and prompt caching cut LLM token cost by 45%.

A self-learning conversational platform that turns clinician questions into SQL against hospital data, serving 125,000+ queries. Backed by a spaCy NER pipeline with an Excel-driven annotation loop so the model retrains on corrected real-world entities.

The simulation and benchmarking engine behind hospital capacity planning — Monte Carlo and discrete-event models, Erlang-C queueing, iterative proportional fitting, and a CMP cost model. Care pathways are modelled as graphs, with DFS and BFS traversal running in production to trace and analyse patient routes through departments.

Built an internal npm package that converts JSON data into formatted Excel files, adopted across projects at Simplogics to cut out repetitive spreadsheet-generation code.

A college event scheduler, manager, and live score-tracking platform built with a team in 16 hours. Placed 1st of 15 teams in the internal Simplogics hackathon.