Vedh Sonawane // Portfolio // Document view
software that doesn't always stay on a screen.
An RC-car chassis rebuilt into an autonomous assistant: Arduino motor control underneath, computer vision on top, and onboard decision-making that closes the loop between what it sees and where it drives. Sensors feed obstacle and target data back into the control layer in real time.
A desk companion that is equal parts firmware and intelligence: embedded C++ on an Arduino Uno drives the hardware and motion, while a Python layer handles the reasoning and automation. The two halves talk over serial, which is the clearest small example of software leaving the screen.
Continuous temperature monitoring drives automatic fan-speed control, with real-time telemetry on an LCD and interactive controls for overriding the curve. A closed control loop you can hold your hand in front of and watch respond.
Two salvaged DVD drive sleds provide the X and Y motion, an old hard drive becomes the base, and a single servo lifts the pen. An Arduino Uno drives the whole thing through an L293D motor shield. Almost every moving part came out of something that was already broken, which is the point: e-waste turned into a machine that draws with precision.
A single command palette that sits over the whole machine: system monitoring, project management, file search, and persistent memory, with sandboxed AI tool-calling so the model can actually act instead of only answering. Runs locally, so nothing leaves the desk.
Give it a high-level goal and it produces a finished outcome, not a plan: it researches, decomposes the goal into steps, and executes end-to-end, reporting what it did at each stage.
Reads behavioural signals only (response timing, participation patterns, thread engagement) and never message content, which is what makes it deployable at all. Surfaces risk weeks before someone hands in notice. Built on the Canvas API, Block Kit and MCP.
Role reversal as a game mechanic: you are the model, the requests keep coming, and the clock is running. Scenarios are AI-generated fresh each run, and you are scored on both speed and judgment, which turns out to be a surprisingly good argument about what these systems are actually doing.
Turns a single prompt into a complete presentation: dynamic themes, generated structure, and real sourced images rather than placeholder boxes. Groq handles inference; FastAPI does the slide assembly.
Tracks what you actually have, warns before it expires, suggests recipes from current inventory, and lets neighbours share surplus instead of binning it.
Multi-step registration with draft autosave on the applicant side; on the organizer side, a dashboard for scoring applications, making accept/reject decisions, and firing automated decision emails. Built for EurekaHacks after competing there.
Small groups make one shared daily commitment. Miss it without spending a wildcard and the whole group's streak resets. The social stake is the product. Google auth, row-level security, realtime sync, and browser-native Web Push fired from a Vercel cron.
Every algorithm implemented from first principles in NumPy, including the classification diagnostics used to evaluate them. Written to understand the maths rather than to call the library.
Generates new fraud patterns that have not appeared in production yet, so detection models see the next attack shape during training instead of after the loss.
Swipe through travel content the way you already do, and Soar assembles the places you liked into an actual routed itinerary with timings, closing the gap between inspiration and a plan.
An AI rewrites the repository's README daily with new lore and a new puzzle. The community solves it through pull requests and issues, and the story branches based on what they find. The repo is the game board.
Analyses high-quality websites and screenshots to work out the patterns underneath them: typography, layout, colour, spacing, visual hierarchy. It keeps that as a Design DNA, uses it to generate or redesign a site, critiques the result visually, and iterates on its own output.
A production-ready full-stack platform built on a modular architecture, with separate agents for fraud detection, health, memory and explainability working together behind one interface. Deliberately general enough to compete across very different hackathon tracks.
Tooling aimed at the WebAssembly attack surface: inspecting and hardening modules that increasingly run untrusted code inside the browser sandbox.
A concept build around hydraulic waste compaction, modelling the mechanism and the control software that would run it.
Reproduces the cadence, hesitation and error-correction of real human typing through PyAutoGUI, instead of the machine-perfect keystroke bursts automation usually produces.
Current conditions and air quality pulled live and rendered for a glance rather than a read.
Submitted
Built Neural Flux.
Built EurekaHACKS Application Portal, Soar.
Built VibeCheck.
Teaches and mentors 100+ students in Python, JavaScript, C#, Unity and MakeCode Arcade, from a first block-code lesson through to working projects.
Founded the school's Python Club: designed the curriculum and taught every level, from absolute beginners to students shipping their own tools. Recognized by the school.
Ran school-wide holiday events for kids aged 5 to 12: scheduling, activity design, and day-of coordination.
Learning Buddies, Wee Walkers and Pizza Helpers, plus ongoing community-centre volunteering.
Transmission 2026.253SIGNAL-LOST @ 3252a70 // 🤖 Day 161 update
Python, TypeScript, JavaScript, C#, C++, Rust, SQL, HTML, CSS, Bash
Next.js, React, Vue, Tailwind, shadcn/ui, Three.js, Node.js, FastAPI, Supabase, PostgreSQL, REST APIs
OpenAI API, Groq, Gemini, MCP, Prompt engineering, LLM agents, OpenCV, NumPy, Core ML from scratch
Git/GitHub, Vercel, AWS, Docker, Figma, Unity, Arduino, Slack, Google Workspace, MS Office
Diffusion models, Transformer architectures, Reinforcement learning