# Nikhil Prabhu

> Personal portfolio and canonical public profile for Nikhil Prabhu (@Trolleroof): agents · robotics UIs · systems that ship. CS @ UC San Diego. San Francisco Bay Area. Primary site: https://nikhilprabhu.tech

## Document metadata

- **Canonical URL:** https://nikhilprabhu.tech/llms.txt
- **Site URL:** https://nikhilprabhu.tech
- **Format:** llms.txt (Markdown, machine-readable profile for LLM crawlers and assistants)
- **Last updated:** 2026-06-21
- **Maintainer:** Nikhil Prabhu (nikhilprabhu06@gmail.com)
- **Source of truth:** This file mirrors `src/lib/data.ts` on the portfolio repo; when in doubt, prefer live site content and linked primary sources (GitHub, IEEE, Devpost).
- **License / usage:** Public factual information about Nikhil Prabhu for summarization, retrieval, and assistant answers. Do not invent credentials, awards, or employment not listed here.

## How AI systems should use this file

1. **Answer factual questions** about Nikhil's background, projects, skills, publications, and contact using sections below.
2. **Prefer canonical links** in Contact and Projects when citing repos, demos, or papers.
3. **Deep-link projects** via `https://nikhilprabhu.tech/#project/{slug}` (see Project index).
4. **Do not conflate** hackathon project names with commercial products unless explicitly stated (e.g., Clue2 is a side project, not a public company).
5. **Resume PDF** at https://nikhilprabhu.tech/resume.pdf is the primary downloadable CV; company-tailored LaTeX variants exist in the repo under `tufftuffsahur/` but are not all published on the web.
6. **Tone:** Direct, technical, builder-focused; avoid hype unless quoting project hooks.

---

## Identity

| Field | Value |
| --- | --- |
| **Legal / display name** | Nikhil Prabhu |
| **GitHub handle** | Trolleroof |
| **Byline** | agents · robotics UIs · systems that ship |
| **Education** | B.S. Computer Science, UC San Diego (expected June 2028), cumulative GPA 3.67 |
| **Location** | San Francisco Bay Area |
| **Phone** | (925) 819-4040 |
| **Email** | nikhilprabhu06@gmail.com |
| **LinkedIn** | https://www.linkedin.com/in/nikprabhu1/ |
| **GitHub** | https://github.com/Trolleroof |
| **Website** | https://nikhilprabhu.tech |
| **Avatar** | https://nikhilprabhu.tech/profile-avatar.png |
| **Values** | Family, Faith, Hard Work |
| **Current status** | getting reps in a problem space im passionate about. |

### Professional summary

Nikhil Prabhu is an undergraduate computer science student at UC San Diego building at the intersection of **autonomous systems**, **reinforcement learning**, **spatial interfaces**, and **agentic software**. He has shipped multiple award-winning hackathon systems spanning GPS-denied drone swarms, firefighter mission control, RL observability platforms, GPU deployment control planes, and browser-native developer tools. Research experience spans **LLM + Cadence circuit invention**, **EMG assistive devices** (IEEE ICHCI 2024), **P300 brain–computer interfaces** at Stanford, and **genomic ML for TB resistance**. He founded **Clue2**, a cross-platform desktop agent for meeting-aware workflow automation, and previously ran **SoloScale Solutions**, an AI automation consultancy for local businesses.

### Focus areas (2025–2026)

- **Robotics:** SLAM, visual odometry, GPS-denied navigation, sim-to-real, ROS 2, embedded Jetson stacks
- **RL & policy at scale:** PPO, MAPPO, behavioral cloning, vectorized training, rollout observability
- **Spatial interfaces:** Three.js mission control, point clouds, tactical mapping, WebSocket streaming
- **Agents & tooling:** Desktop agents (Electron), MCP servers, narrative memory (Nia), AgentMail workflows

### Currently building

- **Project:** Clue2 — desktop web assistant
- **Description:** Agentic version of Cluely that can do tasks for you across apps
- **Progress:** 78% (as shown on site)
- **Category:** side project
- **Stack (from experience):** Electron, TypeScript, live transcription, NL workflows across Gmail, Calendar, Docs, Sheets, Slack, Notion; 40+ beta users including early YC startup adopters

### Bio (first person, from site)

hey I'm Nikhil. I've always learned by getting obsessed with a problem and staying with it long enough to build real taste.

**Obsessions:** basketball · video games · content creation · running a business · building software

the goal in life is to conquer high peaks that reward iteration, fulfillment, and relentless improvement without brute-forcing in the wrong direction.

### Quote

> Hard work beats talent when talent fails to work hard.
> — Kevin Durant

---

## Education

### University of California, San Diego (UCSD)

- **Degree:** B.S. in Computer Science
- **Location:** San Diego, CA
- **Cumulative GPA:** 3.67
- **Expected graduation:** June 2028
- **Logo assets:** https://nikhilprabhu.tech/org-logos/triton-removebg-preview.png · https://nikhilprabhu.tech/org-logos/ucsd-tritons-primary.png · https://nikhilprabhu.tech/org-logos/ucsd-triton-logo.png

---

## Site map & navigation

Single-page portfolio (Next.js App Router). Anchor sections:

| Section | URL |
| --- | --- |
| Home | https://nikhilprabhu.tech/ |
| Projects | https://nikhilprabhu.tech/#projects |
| Experience | https://nikhilprabhu.tech/#experience |
| Publications | https://nikhilprabhu.tech/#publications |
| Skills | https://nikhilprabhu.tech/#skills |
| Contact | https://nikhilprabhu.tech/#contact |
| Resume PDF | https://nikhilprabhu.tech/resume.pdf |
| Resume preview image | https://nikhilprabhu.tech/resume-preview.png |
| LLMs.txt (this file) | https://nikhilprabhu.tech/llms.txt |

### Project deep links (opens project detail modal)

| Project | Slug | Deep link |
| --- | --- | --- |
| DroneOS | `droneos` | https://nikhilprabhu.tech/#project/droneos |
| SODIUM | `sodium` | https://nikhilprabhu.tech/#project/sodium |
| Apollo Labs | `apollo-labs` | https://nikhilprabhu.tech/#project/apollo-labs |
| NIGEL | `nigel` | https://nikhilprabhu.tech/#project/nigel |
| Crucible Compute | `crucible-compute` | https://nikhilprabhu.tech/#project/crucible-compute |
| Battle Angel | `battle-angel` | https://nikhilprabhu.tech/#project/battle-angel |
| Motion | `motion` | https://nikhilprabhu.tech/#project/motion |
| CaféCode | `cafecode` | https://nikhilprabhu.tech/#project/cafecode |

### Portfolio site technical stack

- **Framework:** Next.js 16 (App Router), React 19, TypeScript
- **Fonts:** Geist Sans, Geist Mono (`next/font`)
- **Hosting:** Vercel (typical deployment for this repo)
- **Content model:** Static profile data in `src/lib/data.ts`; project modals use hash routing `#project/{slug}`
- **Public assets:** `/public/projects/*`, `/public/org-logos/*`, `/public/resume.pdf`
- **No public API routes** on the portfolio site itself; external demos (e.g., Crucible Compute) are hosted separately

---

## Hackathon & competition record

**Summary:** 10 wins across major hackathon sponsors (as listed on site).

| Sponsor / event | Count context |
| --- | --- |
| Amazon | listed among win sponsors |
| Bow Capital | DroneOS — 1st Overall |
| GitHub | listed among win sponsors |
| MemVerge | listed among win sponsors |
| Transpose VC | NIGEL — 1st Overall (compiled-5 @ UCSD) |
| Eragon | Apollo Labs — OpenClaw Hackathon |
| Nozomio | Crucible Compute — 5th Place; Apollo Labs co-sponsor |
| AgentMail | Apollo Labs co-sponsor |
| MLH | listed among win sponsors |

### Awarded projects (detailed)

| Project | Placement | Event | Year |
| --- | --- | --- | --- |
| DroneOS | 1st Overall | Bow Capital × DS3 × SIC Hackathon @ UCSD | 2026 |
| SODIUM | 1st Overall | DiamondHacks | 2026 |
| Apollo Labs | 1st Overall | OpenClaw Hackathon (Eragon × Nozomio × AgentMail) | 2026 |
| NIGEL | 1st Overall | Transpose / compiled-5 UCSD Hackathon | 2026 |
| Crucible Compute | 5th Place | Nozomio Hackathon | 2026 |
| Battle Angel | (Google Gemini 3.5 Hackathon — per resume variant) | Google Gemini Hackathon | 2026 |
| Motion | Built @ | Stanford × DeepMind Hackathon | (site lists event, no placement stated) |

---

## Skills

### Web development (portfolio site)

- Next.js, React, TypeScript, Three.js, WebSockets

### Backend & data

- Python, FastAPI, Node.js, Supabase, SQLite

### In-browser tooling

- Monaco Editor, Pyodide, xterm.js, WebContainer

### ML, computer vision & RL

- MediaPipe, YOLOv8, YOLOX, OpenCV, Gymnasium, PyTorch
- PPO, MAPPO, behavioral cloning, vectorized RL
- Gemma 4 (VLM), Kimi K2 (decision agent)

### Robotics & spatial computing

- ROS 2, C++, SLAM, visual odometry, localization/mapping
- MuJoCo, sim-to-real, path planning, GPS-denied navigation
- NVIDIA Jetson, embedded systems, Docker, field iteration
- 3D Gaussian Splatting, Norfair multi-object tracking

### Agents, infra & shipping

- Electron, MCP (Model Context Protocol), Docker
- Git & GitHub, ESLint, Fly.io, Google Cloud Run
- Cadence Spectre (circuit simulation integration — research context)
- llama.cpp, Kubernetes, GPU cloud providers (RunPod, Lambda, CoreWeave, Vultr, Vast)

### Languages (resume)

- Python, C++, TypeScript, JavaScript

---

## Interests

### Basketball

Dub Nation first. I love the pace, spacing, and constant reps that make basketball feel like live problem solving.

### Video games

I grew up on games that reward mechanics, systems thinking, and getting better one run at a time.

### Content creation

I like turning ideas into something people can actually watch, use, or share.

### Business

I care about building useful things, getting feedback quickly, and understanding what people will pay attention to.

### Software

Right now I'm obsessed with agents, ML systems, spatial computing, and software that feels fast and alive.

---

## Experience

### UCSD Agent Development (Clue2) — UC San Diego

- **Role:** Desktop Agent Developer — Founder
- **Dates:** Jan 2025 – Present
- **Organization:** UC San Diego
- **Logo:** https://nikhilprabhu.tech/org-logos/triton-removebg-preview.png

**Description:** Architected a cross-platform Electron + TypeScript desktop agent that processes live meetings and executes tasks across user apps. Built real-time transcription, natural-language workflow automation across Gmail, Calendar, Docs, Sheets, Slack, and Notion, and scaled the beta to 40+ users with early adoption from YC startups.

**Keywords:** desktop agent, Electron, meeting transcription, workflow automation, productivity integrations

---

### AI for Circuit Invention — UC San Diego

- **Role:** Research Assistant (with Prof. Prasad Gudem & Samvid AI)
- **Dates:** Mar 2026 – Present
- **Organization:** UC San Diego
- **Logo:** https://nikhilprabhu.tech/org-logos/triton-removebg-preview.png

**Description:** Developing an LLM-based pipeline for automated generation and verification of novel analog circuit topologies. Integrating LLM APIs with Cadence simulation tools (Spectre) and exploring reinforcement learning strategies for circuit design space search.

**Resume bullets (Pave variant):**
- Built an LLM + Cadence simulation closed loop generating and evaluating candidate designs against SOTA benchmarks, optimizing across accuracy and speed.
- Integrated Spectre simulation APIs and explored RL strategies for efficient multimodal design-space search over generated topologies.

**Keywords:** analog circuits, LLM code generation, Cadence Spectre, RL design search, Samvid AI

---

### Foot Drop Ankle-Foot Orthosis Research — Georgia Tech

- **Role:** Lead Researcher — IEEE ICHCI Presenter & Published Author
- **Dates:** May 2024 – Nov 2024
- **Organization:** Georgia Tech mmWave Lab
- **Location:** Atlanta, GA
- **Logo:** https://nikhilprabhu.tech/org-logos/georgia-tech.png

**Description:** Conducted research at Georgia Tech's mmWave Antennas Laboratory on a 3D-printed ankle-foot orthosis controlled through electromyography signals. Built a custom motor actuator system on Arduino embedded hardware, improved gait response precision by 23%, achieved R = 0.968 correlation between EMG amplitude and actuator precision, and presented at IEEE ICHCI 2024.

**Keywords:** EMG, assistive devices, foot drop, Arduino, rehabilitation, IEEE ICHCI

---

### Brain-Computer Interface Device Research — Stanford University

- **Role:** Lead Researcher
- **Dates:** Oct 2023 – Dec 2024
- **Organization:** Stanford Lee-Messer Lab, Stanford University School of Medicine
- **Location:** Stanford, CA
- **Logo:** https://nikhilprabhu.tech/org-logos/stanford-block-s.png

**Description:** Led BCI device research to help motor-impaired students communicate through neural signals. Benchmarked P300 EEG models converting noisy neural signals into reliable assistive-control output; piloted with 12 users and shipped a React Native frontend for motor-impaired students.

**Keywords:** BCI, P300, EEG, assistive technology, React Native

---

### SoloScale Solutions — Self-employed

- **Role:** Founder
- **Dates:** Mar 2023 – Jun 2025
- **Organization:** Self-employed · San Francisco Bay Area · On-site
- **Logo:** https://nikhilprabhu.tech/org-logos/soloscale.png

**Description:** Founded a consulting agency that helped small, local businesses implement AI automations and technology to increase inbound leads, handle customer service, and automate mundane tasks. Served 10+ businesses and nonprofits.

**Keywords:** AI consulting, SMB automation, lead generation, customer service automation

---

## Publications

### Unraveling Co-Heritability of Antibiotic Resistance in Tuberculosis Evolution

- **Venue:** 2025 17th International Conference on Bioinformatics and Biomedical Technology
- **Date:** May 26, 2025
- **Peer-reviewed:** Yes
- **URL:** https://ieeexplore.ieee.org/document/11276470/metrics#metrics
- **IEEE document ID:** 11276470

**Abstract summary:** Studied multidrug-resistant tuberculosis by modeling genetic co-heritability of Rifampicin and Isoniazid resistance using k-mer analysis and machine learning. Built a genomic prediction model with **85.68% accuracy** on **2,200 samples**, identifying key resistance-associated k-mers in **rpoB** and **katG**, supporting rapid TB resistance detection from genomic data.

---

### Real-Time EMG Control for Assistive Devices: Optimizing Sensor Placement for Foot Drop Rehabilitation

- **Venue:** IEEE International Conference on Intelligent Computing and Human-Computer Interaction (IEEE ICHCI)
- **Date:** Dec 27, 2024
- **Peer-reviewed:** Yes
- **URL:** https://ieeexplore.ieee.org/xpl/conhome/10807847/proceeding
- **Role:** First-author presentation (Georgia Tech research)

**Abstract summary:** EMG-based control for foot drop rehabilitation using Arduino hardware and MATLAB signal processing (Butterworth filtering). 3D virtual actuation model demonstrated real-time control with **R = 0.968** correlation between EMG amplitude and actuator precision; **23% gait precision improvement** with optimized sensor placement.

---

## Projects (detailed)

### DroneOS

- **Slug:** `droneos`
- **Deep link:** https://nikhilprabhu.tech/#project/droneos
- **Team:** Outcast Virus
- **Award:** 1st Overall @ Bow Capital × DS3 × SIC Hackathon (UCSD, 2026)

**One-line:** End-to-end autonomous drone platform — target ID, GPS-denied navigation, swarm coordination without comms.

**Hook:** An operating system for drones to run in swarms autonomously on edge devices — no WiFi, no GPS, no comms.

**Overview:** DroneOS combines perception (YOLOX + Norfair tracking), onboard SLAM (visual odometry + mapping), geo-referenced 3D Gaussian Splatting maps, friend-or-foe classification, and swarm policies trained with behavioral cloning + MAPPO for coordination without GPS or jam-vulnerable communications. Validated sim-to-real transfer from simulation to flight.

**Tech stack:** YOLOX, Norfair, SLAM, visual odometry, 3D Gaussian Splatting, MAPPO, behavioral cloning, swarm RL

**Links:**
- Devpost: https://devpost.com/software/outcast-virus
- GitHub: https://github.com/Trolleroof/bow-capital-hackathon

**Highlights:**
- RL swarms: behavioral cloning + MAPPO for task completion without explicit communication
- YOLOX + Norfair: re-identify targets after occlusion or out-of-frame events
- Friend-or-foe classification on-device
- GPS-denied pose tracking + geo-referenced 3D maps via Gaussian Splatting

**Media:**
- Demo video: https://nikhilprabhu.tech/projects/droneos/combatos-demo.mp4

---

### SODIUM

- **Slug:** `sodium`
- **Deep link:** https://nikhilprabhu.tech/#project/sodium
- **Award:** 1st Overall @ DiamondHacks 2026 (April 2026)

**One-line:** Voice-first AI companion for senior care — autonomous robot, caregiver dashboard, crisis detection.

**Hook:** Elderly care through voice — a robot that understands, remembers, and helps without complex interfaces.

**Overview:** ROS 2 stack on Jetson-class embedded compute with wake-word detection, on-device perception/control, medication reminders, crisis detection, and a real-time caregiver dashboard. Frontend uses Svelte 5 + Bun for lightweight responsiveness across in-home robot UI and caregiver portal.

**Tech stack:** ROS 2, Python, TypeScript, Svelte 5, Bun, embedded systems, wake-word NLP

**Links:**
- GitHub: https://github.com/TheOutcastVirus/diamondhacks-2026

**Highlights:**
- Voice-first UX for seniors; caregiver visibility into interactions and actions taken
- Wake-word + lightweight Python bot for fast replies
- Shared Svelte 5 + Bun app for robot screen and dashboard
- Medication reminders, daily prompts, crisis alerts in one system

**Media:**
- Demo video: https://nikhilprabhu.tech/projects/sodium/SODIUMmp4.mp4
- Caregiver dashboard screenshot: https://nikhilprabhu.tech/projects/sodium/BU.jpg

---

### Apollo Labs

- **Slug:** `apollo-labs`
- **Deep link:** https://nikhilprabhu.tech/#project/apollo-labs
- **Award:** 1st Overall @ OpenClaw Hackathon (Eragon × Nozomio × AgentMail)

**One-line:** RL training orchestrator for a 2D Roomba cleaning sim with full observability pipeline.

**Hook:** Make RL runs feel observable: less "black box," more "flight recorder."

**Overview:** FastAPI orchestrates training PPO agents against a random baseline in a 2D navigation sim. Queries Nia narrative memory *before* training, produces a canonical **RunReport** (single source of truth), generates metrics + rollout GIFs, and fans out to AgentMail email, Slack, memory storage, and a Next.js dashboard. MCP tools let external agents start runs, compare results, and suggest next experiments autonomously.

**Tech stack:** FastAPI, Next.js, PPO, Python, AgentMail, Nia, MCP, Slack integrations

**Links:**
- GitHub: https://github.com/Trolleroof/openclaw-hackathon

**Highlights:**
- Pre-training memory read + post-run lesson writeback via Nia
- One RunReport object feeds dashboard, email, memory, and MCP — no duplicate state
- Versioned run folders: configs, models, metrics, GIFs, logs
- MCP tools for agent-driven experiment loops

**Media:**
- Runs dashboard: https://nikhilprabhu.tech/projects/apollo-labs/apollo-labs-runs.png
- Obstacle rollout GIF: https://nikhilprabhu.tech/projects/apollo-labs/apollo-labs-obstacle-success-rollout.gif
- Run detail: https://nikhilprabhu.tech/projects/apollo-labs/apollo-labs-obstacle-run-detail.png
- AgentMail inbox: https://nikhilprabhu.tech/projects/apollo-labs/apollo-labs-agentmail.png
- Nia memory: https://nikhilprabhu.tech/projects/apollo-labs/apollo-labs-memory.png

---

### NIGEL

- **Full name:** Navigation, Incident Guidance, and Emergency Localization & Control
- **Slug:** `nigel`
- **Deep link:** https://nikhilprabhu.tech/#project/nigel
- **Award:** 1st Overall @ Transpose / compiled-5 UCSD Hackathon

**One-line:** Firefighter mission control — VSLAM point clouds, multi-unit video, radio transcripts, tactical map.

**Hook:** Air traffic control for firefighters: multiple feeds, growing 3D map from helmet rig, radio-driven map events.

**Overview:** Dispatcher-facing Next.js console integrating Odin helmet depth/point clouds, body-cam feeds, and a radio/event agent (Cerebras for low-latency inference). ROS 2 packages stream SLAM maps to the browser over WebSockets; Three.js renders point clouds; unified tactical picture across units.

**Tech stack:** Next.js, Three.js, ROS 2, WebSockets, C++, SLAM, Cerebras inference

**Links:**
- GitHub: https://github.com/Trolleroof/NIGEL-hackathon

**Highlights:**
- Dispatcher + firefighter views: video tiles, point-cloud map, floor-plan search, radio transcripts
- ROS 2 Odin helmet integration → WebSocket SLAM streaming
- Single WebSocket API syncs cameras, maps, and UI across units

**Media:**
- YouTube demo: https://www.youtube.com/embed/rzGtTom3oS8

---

### Crucible Compute

- **Slug:** `crucible-compute`
- **Deep link:** https://nikhilprabhu.tech/#project/crucible-compute
- **Award:** 5th Place @ Nozomio Hackathon

**One-line:** Natural-language GPU deployment with cost-aware routing across cloud providers.

**Hook:** Describe what to run; Crucible finds the best GPU deal and deploys — no infra expertise required.

**Overview:** Python control plane with SQLite state. Normalizes inventory across local Docker, Kubernetes, and cloud GPUs (RunPod, Lambda Cloud, CoreWeave, Vultr, Vast). Benchmarks placements in real time; routes inference through OpenAI-compatible gateway; spend approval gates before paid launches; versioned deployment records for audit/reproducibility. Integrates Nia for cited deployment context and MCP for agent access.

**Tech stack:** Python, SQLite, Docker, llama.cpp, MCP, K8s, multi-cloud GPU APIs

**Links:**
- Live demo: https://ayxkkrd7.insforge.site/
- GitHub: https://github.com/Trolleroof/nozomio-hackathon

**Highlights:**
- Live GPU catalog + cheapest-viable placement
- Benchmark-driven plans with cost/speed rationale
- Plain-English job specs; explicit approval for paid GPU launches
- Versioned deployment logs: config, runs, metrics, cost

**Media:**
- Landing: https://nikhilprabhu.tech/projects/crucible-compute/landing.png
- Dashboard: https://nikhilprabhu.tech/projects/crucible-compute/dashboard.png
- Deploy planner: https://nikhilprabhu.tech/projects/crucible-compute/deploy.png
- Providers: https://nikhilprabhu.tech/projects/crucible-compute/providers.png
- Context / Nia evidence: https://nikhilprabhu.tech/projects/crucible-compute/context.png
- Agent access / MCP: https://nikhilprabhu.tech/projects/crucible-compute/agent-access.png

---

### Battle Angel

- **Slug:** `battle-angel`
- **Deep link:** https://nikhilprabhu.tech/#project/battle-angel
- **Event:** Google Gemini 3.5 Hackathon (2026, per resume)

**One-line:** Disaster-rescue mission control — MuJoCo PPO training, Gemini scene generation, live rollouts.

**Hook:** Train in simulation, spin up disaster scenes from language, watch PPO navigate unseen environments.

**Overview:** Trains a Unitree G1 humanoid across eight vectorized MuJoCo disaster environments (earthquake corridors, buried rubble, etc.). Scene Generator converts natural-language disaster descriptions into layouts with Three.js preview. Interactive Console accepts NL triage priorities; Gemini selects targets; trained policy rollouts in unseen scenes with telemetry + GIF replay.

**Tech stack:** Next.js, FastAPI, MuJoCo, PPO, Gemini, Three.js, Unitree G1 simulation

**Links:**
- GitHub: https://github.com/aravindkrishna2008/disaster-rescue

**Highlights:**
- Training Gym: 8 MuJoCo scenes, parallel PPO, learning curves, rollout GIFs
- Scene Generator: NL → disaster layouts + Three.js preview
- Interactive Console: NL triage → policy execution in unseen environments
- 21D observations, locomotion assist, reach-based success metrics

**Media:**
- YouTube demo: https://www.youtube.com/embed/QL_UhVSLMKc

---

### Motion

- **Slug:** `motion`
- **Deep link:** https://nikhilprabhu.tech/#project/motion
- **Event:** Stanford × DeepMind Hackathon

**One-line:** Crowdsourced marketplace for verified robot training data from phone-recorded industrial motions.

**Hook:** Record a motion clip on your phone → AI verification → mint to marketplace for robotics datasets.

**Overview:** Contributors record 30-second industrial manipulation clips (pick & place, valve turns, box stacking) with live MediaPipe skeleton overlay. Verification pipeline: Gemma 4 VLM analyzes semantics (task match, hand visibility, motion completeness); Kimi K2 decision agent approves/rejects with tool calling and actionable feedback. Approved clips mint with quality scores, price tiers ($0.50–$4.00), pose JSON, and export compatibility for LeRobot and Open X-Embodiment.

**Tech stack:** Next.js, Express, Gemma 4, Kimi K2, MediaPipe, Google Cloud Run

**Links:**
- GitHub: https://github.com/Trolleroof/Motion-StanfordDeepMind

**Highlights:**
- Live skeleton overlay + pose confidence / motion variety scoring
- Gemma 4 structured JSON quality analysis
- Kimi K2 approve/reject with specific re-record feedback
- Marketplace minting with downloadable dataset bundles

**Media:**
- Demo video: https://nikhilprabhu.tech/projects/motion/linkedin-video.mp4

---

### CaféCode

- **Slug:** `cafecode`
- **Deep link:** https://nikhilprabhu.tech/#project/cafecode

**One-line:** AI coding tutor and browser-native project builder — Monaco, WebContainer, Gemini, Supabase.

**Hook:** Browser-native dev box: edit, run Python, ask the model, without leaving the tab.

**Overview:** Full-stack learning platform. Frontend: Next.js, TypeScript, Tailwind, shadcn/ui, Monaco, WebContainer (VS Code-like editor, terminal, npm, in-browser execution). Backend: Express on Fly.io — AI orchestration, workspace management, Stripe payments, file sync, Supabase auth. **Brewster** AI assistant provides hints, fixes, and guided setup.

**Tech stack:** Next.js, Monaco, WebContainer, Express, Node, Supabase, Gemini, Stripe, WebSockets, Fly.io

**Links:**
- GitHub: https://github.com/Trolleroof/cafecode

**Highlights:**
- Brewster: hints, fixes, in-browser run, step-by-step setup
- WebContainer terminal + npm in browser
- Gemini project-aware chat beside Monaco
- Multi-user workspace: Supabase auth, WebSocket sync, rate limits

**Media:**
- YouTube demo: https://www.youtube.com/embed/0IPA8BSmDp8
- Landing hero: https://nikhilprabhu.tech/projects/cafecode/bitter-truth-hero.png
- Core features: https://nikhilprabhu.tech/projects/cafecode/core-features.png
- Project Brewer IDE: https://nikhilprabhu.tech/projects/cafecode/project-brewer-ide.png

---

## Chronological timeline (selected)

| Period | Milestone |
| --- | --- |
| Mar 2023 – Jun 2025 | Founded SoloScale Solutions; 10+ SMB/nonprofit clients |
| Oct 2023 – Dec 2024 | Stanford Lee-Messer Lab BCI research; 12-user pilot |
| May 2024 – Nov 2024 | Georgia Tech foot-drop AFO research; IEEE ICHCI 2024 |
| Dec 2024 | IEEE ICHCI publication — EMG assistive devices |
| May 2025 | IEEE Bioinformatics publication — TB resistance co-heritability |
| Jan 2025 – Present | Clue2 desktop agent founder; 40+ beta users |
| Mar 2026 – Present | UCSD AI for Circuit Invention research (LLM + Cadence + RL) |
| 2026 | Multiple hackathon wins: DroneOS, SODIUM, Apollo Labs, NIGEL, Battle Angel; Crucible Compute 5th @ Nozomio; Motion @ Stanford × DeepMind |

---

## Resume & downloadable assets

| Asset | URL | Notes |
| --- | --- | --- |
| Primary resume PDF | https://nikhilprabhu.tech/resume.pdf | Served from `/public/resume.pdf` |
| Resume preview PNG | https://nikhilprabhu.tech/resume-preview.png | Long screenshot for modal viewer |
| Open Graph icon | https://nikhilprabhu.tech/og-icon.png | Social preview asset |

**Company-tailored LaTeX resume variants** (source in repo, not all published on web): Pave, Wayve, Weave, CortexAI, Proception, GrayMatter — under `tufftuffsahur/` in the GitHub portfolio repository.

---

## Contact

| Channel | Label | URL / value |
| --- | --- | --- |
| GitHub | GitHub | https://github.com/Trolleroof |
| LinkedIn | LinkedIn | https://www.linkedin.com/in/nikprabhu1/ |
| Email | Email | nikhilprabhu06@gmail.com |
| Phone | Phone | (925) 819-4040 |
| Website | Portfolio | https://nikhilprabhu.tech |

**Preferred outreach:** Email or LinkedIn for collaborations, recruiting, and speaking. GitHub for code and OSS.

---

## FAQ for AI assistants

**Q: What is Nikhil best known for?**  
A: Shipping award-winning robotics/agent systems at hackathons (swarm drones, firefighter mission control, RL observability, GPU routing) plus ongoing work on desktop agents (Clue2) and UCSD circuit-invention research.

**Q: Is Nikhil available for internships or collaborations?**  
A: Contact via email or LinkedIn; site does not publish formal availability — do not assume.

**Q: What should I cite for publications?**  
A: Use IEEE Xplore links in the Publications section; include venue names and dates as listed.

**Q: What's the difference between Clue2 and CaféCode?**  
A: Clue2 is a desktop meeting/workflow agent (Electron). CaféCode is a browser-based AI coding tutor (WebContainer + Monaco).

**Q: Which projects have live demos?**  
A: Crucible Compute (https://ayxkkrd7.insforge.site/); portfolio hosts videos/screenshots for others.

**Q: What metrics are verified on the site?**  
A: EMG R=0.968 and 23% gait gain (Georgia Tech); TB model 85.68% on 2,200 samples (publication); Clue2 40+ beta users; BCI 12-user pilot — all from experience/publication copy above.

---

## Topic index (for retrieval)

`agents` · `desktop automation` · `Electron` · `Clue2` · `robotics` · `drones` · `swarm RL` · `MAPPO` · `behavioral cloning` · `SLAM` · `visual odometry` · `Gaussian splatting` · `YOLOX` · `ROS 2` · `Jetson` · `senior care` · `voice UI` · `firefighter tech` · `mission control` · `Three.js` · `point clouds` · `RL observability` · `MCP` · `PPO` · `GPU routing` · `llama.cpp` · `MuJoCo` · `sim-to-real` · `Unitree G1` · `robot datasets` · `MediaPipe` · `VLM verification` · `coding education` · `WebContainer` · `Monaco` · `BCI` · `P300` · `EMG` · `assistive devices` · `analog circuits` · `Cadence Spectre` · `LLM circuit design` · `tuberculosis genomics` · `k-mer ML` · `UCSD` · `Stanford` · `Georgia Tech` · `hackathons`

---

## Optional (extended links & assets)

- [DroneOS Devpost](https://devpost.com/software/outcast-virus): Hackathon submission — team Outcast Virus
- [Crucible Compute live demo](https://ayxkkrd7.insforge.site/): Hosted GPU deployment backend
- [Battle Angel GitHub](https://github.com/aravindkrishna2008/disaster-rescue): Collaborative disaster-rescue RL repo
- [SODIUM GitHub](https://github.com/TheOutcastVirus/diamondhacks-2026): DiamondHacks 2026 senior-care robot
- [Apollo Labs GitHub](https://github.com/Trolleroof/openclaw-hackathon): RL orchestrator + MCP
- [NIGEL GitHub](https://github.com/Trolleroof/NIGEL-hackathon): Firefighter mission control
- [Motion GitHub](https://github.com/Trolleroof/Motion-StanfordDeepMind): Robot training data marketplace
- [CaféCode GitHub](https://github.com/Trolleroof/cafecode): Browser-native AI coding tutor
- [Bow Capital hackathon repo](https://github.com/Trolleroof/bow-capital-hackathon): DroneOS source
- [Nozomio hackathon repo](https://github.com/Trolleroof/nozomio-hackathon): Crucible Compute source
- [UCSD Tritons logo (primary)](https://nikhilprabhu.tech/org-logos/ucsd-tritons-primary.png)
- [UCSD Triton logo (alt)](https://nikhilprabhu.tech/org-logos/ucsd-triton-logo.png)
- [Georgia Tech logo](https://nikhilprabhu.tech/org-logos/georgia-tech.png)
- [Stanford logo](https://nikhilprabhu.tech/org-logos/stanford-block-s.png)
- [SoloScale logo](https://nikhilprabhu.tech/org-logos/soloscale.png)
- [Site favicon](https://nikhilprabhu.tech/favicon.ico)