Work — Macha Praveen

Selected work — AI agents, MCP servers, medical imaging, and computer-vision pipelines shipped end-to-end.

lanforge-mcp

MCP Server · Wi-Fi Test Automation

Python · MCP SDK · httpx · FastMCP

2025

Built an MCP server that lets Claude (or any MCP client) drive Candela Technologies' LANforge Wi-Fi test platform over its JSON-over-HTTP API.
Six tools — list ports, create stations, start L3 traffic, get results, get events, diagnose failures — with mutating tools defaulting to a safe dry-run.
Ships with a mock mode (recorded JSON fixtures) so it runs and demos with zero hardware.
63 tests, 88% coverage, mypy --strict clean; Pydantic v2 schemas for every tool.

Loading post…

AI Glaucoma Screening from Retinal Images

TensorFlow · Keras · U-Net · Flask · React

2025

Built a full-stack medical imaging app that screens retinal fundus images for glaucoma using a dual deep-learning pipeline.
Combined a classification model with a U-Net segmentation model to detect disease and delineate the optic cup and disc.
Automatically computes the Cup-to-Disc Ratio (CDR) and grades severity, with side-by-side mask visualizations.
Flask REST API backend with an interactive React frontend for image upload and cropping.

Conversational AI Tutor

Tavus API · GPT-4o · Flask

2024

Built two AI tutoring modes: a live video-avatar conversation and on-demand generated video answers.
Uses Tavus personas/replicas for a consistent avatar; GPT-4o generates step-by-step teaching scripts.
Live mode embeds real-time conversations; simple mode polls Tavus for rendered video.
Contextual awareness lets the tutor adapt to the current topic.

Computer Vision · Construction

PyTorch 2.x · Faster R-CNN · FastAPI

2024

Built a computer-vision system that detects and counts rebar in construction images, with a React upload + live-camera UI.
Trained a Faster R-CNN detector in PyTorch 2.x for accurate object localisation and counting on real jobsite photos.
Added Indian-market pricing logic — diameter, weight, and cost estimation baked into the detection output.
Served predictions over a FastAPI endpoint and exports full reports as PDF or Excel.

Home

Analyze

Results