Vaibhav Kumar

Vaibhav Kumar

 



 

 

 

---

About

I'm building MECH — an AI and robotics startup. Day to day: training language models on custom datasets, writing real-time computer vision pipelines, shipping desktop software, and prototyping embedded hardware.

The technical direction is toward systems where AI reasoning connects directly to physical actuation. Companion AI, B2B SaaS, and desktop tools are the near-term products. Autonomous hardware is where the work is heading.

Currently focused on Mio, an AI companion fine-tuned on a 1M+ conversation dataset, and Reception, an AI receptionist platform for offices. Learning Rust and ROS2 for the infrastructure and robotics work ahead.

---

MECH

An AI & Robotics startup. Every product contributes to a single long-term ecosystem —software intelligence connected to physical hardware, where each layer makes the next one possible.

Software Products

| Product | Category | What it solves | Status |
|:--------|:---------|:---------------|:------:|
| Mio | AI Companion | Conversational AI with genuine long-term context. Fine-tuned on 1M+ conversation pairs, local LLM, no cloud dependency | Active |
| Reception | B2B SaaS | AI receptionist — visitor management, appointment scheduling, natural conversation. India and international markets | Active |
| AetherXD | Desktop Platform | Local agent runtime and automation shell. AI on your machine, not routed through the cloud | Active |
| Revolver D | Content Pipeline | AI-assisted video production and design automation. Reduces iteration time for independent creators | Active |
| DocaMech | Utility | Automatic file organizer. Monitors Downloads, auto-sorts by type, ships as a licensed Windows application | Prototype |
| Doraemon | Personal AI | Persistent memory, task automation, system control. Infrastructure for an AI that genuinely knows you | Planned |

Hardware Vision

| System | Description | Horizon |
|:-------|:------------|:-------:|
| Robot Arm | AI-controlled mechanical actuation driven by the MECH software stack | Future |
| MECH Core | Shared IoT firmware: Tank Level Checker · Auto Motor Controller · Leak Detector · Environment Monitor | Planned |
| Rail System | Automated workspace equipment transport via rail-mounted actuators | Future |
| Smart Projector | AI-driven dynamic information surfaces in physical space | Concept |
| VR Workspace | Spatial AI environment — productivity tooling in an immersive workspace | Concept |

---

Current Focus

Building

| Project | Stage |
|:--------|:------|
| Mio v1 | Dataset pipeline complete → LoRA fine-tune on 1M+ corpus → local inference deployment |
| Reception | Feature-complete MVP → pricing → India go-to-market |
| AetherXD | Agent runtime core → desktop shell → local model orchestration layer |
| MECH infrastructure | Cross-product shared APIs, data contracts, and reusable primitive libraries |

Learning

| Subject | Why |
|:--------|:----|
| Rust | Systems programming for performance-critical MECH components and embedded targets |
| ROS2 | Robotics middleware — foundation for integrating the MECH hardware stack |
| Deep Learning | Formal study via fast.ai, deeplearning.ai, and building architectures from scratch |
| Autonomous Systems | SLAM, motion planning, sensor fusion, and control theory |
| Embedded Systems | Low-level hardware programming toward the MECH IoT platform |

Research

| Area | Direction |
|:-----|:----------|
| Local inference | Quantization, model compression, and edge deployment without cloud fallback |
| Embodied AI | Connecting perception and reasoning pipelines to physical actuation |
| Agent architectures | Multi-agent coordination over shared memory and task infrastructure |

---

Technology

Languages

Python  ·  TypeScript  ·  JavaScript  ·  C++  ·  C#  ·  Rust (learning)

AI / ML

PyTorch  ·  TensorFlow  ·  Hugging Face Transformers  ·  LangChain  ·  Ollama  ·  LoRA fine-tuning  ·  RAG pipelines  ·  local LLM deployment  ·  model quantization  ·  Gemini API  ·  OpenAI API

Computer Vision

OpenCV  ·  MediaPipe  ·  pose estimation  ·  hand tracking  ·  real-time detection  ·  Flask model serving

Backend

Frontend

Desktop

3D & Simulation

Infrastructure

Databases

Hardware & Embedded

Arduino  ·  Raspberry Pi  ·  ROS (concepts)  ·  ROS2 (learning)  ·  sensor integration  ·  embedded control

---

Projects

> MECH products are listed above. Below are research builds, engineering experiments, and applied work.

Computer Vision

emotion-detection-project  —  Real-time facial emotion recognition. Built to run as a deployable service, not just a demo.
→ View  ·  OpenCV · MediaPipe · Flask

rain-detect  —  Computer vision utility for detecting weather conditions from live camera feeds.
→ View  ·  OpenCV · Python

Local AI Research

RAG-for-IBM  —  Retrieval-augmented generation pipeline for document Q&A. Built under competition constraints.
→ View  ·  LangChain · embeddings · vector search

Basic_ai_from_scratch  —  Core deep learning primitives without frameworks — forward pass, backpropagation, loss functions. Study project running alongside formal Deep Learning coursework.
→ View  ·  Python · NumPy

Dataset Engineering

Dataset Toolkit  —  Internal pipeline for cleaning, normalizing, and structuring large text corpora at scale. Handles 1M+ row datasets. Became the backbone of Mio's training data — subtitle cleaning, CSV repair, language filtering, JSON conversion, batch text normalization.
Private  ·  Pandas · Python · multiprocessing

Robotics & Embedded

Robotics & Embedded R&D  —  Experiments in embedded control, sensor integration, ROS architecture, and Unity-based simulation. Foundation work for the MECH hardware roadmap.
Arduino · Raspberry Pi · ROS · Unity

Local LLM Lab

Benchmarking environment for Qwen, Mistral, Zephyr, Gemini, and Ollama-hosted models. Evaluates RAM footprint, inference speed, and output quality to make integration decisions across MECH products.
Ollama · Hugging Face · quantization testing

VTuber / AI Avatar Research

Real-time avatar pipeline: VRM model → VSeeFace → MediaPipe body tracking → OBS output. Research into AI-driven character animation and streaming presence.
VRM · VSeeFace · MediaPipe · Unity · Blender

Applied & Shipped

Project-sentinal  —  Full-stack crime pattern analysis platform. Crime heatmaps, AI assistant, Docker deployment. Built for the H2S datathon.
→ View  ·  FastAPI · React · Docker

mindmendor  —  Mental health assistant chatbot.
→ View  ·  TypeScript · Vercel

weather-checking  —  Deployed weather utility application.
→ View  ·  TypeScript · Vercel

---

Research Interests

Local AI Inference  —  Cloud dependency is a design flaw in most AI products. Model compression, quantization, and edge deployment are engineering problems worth solving well. Most of what MECH ships needs to work offline and on consumer hardware.

Embodied AI  —  The gap between a model that reasons well and a robot that acts correctly is still enormous. I'm interested in the perception-to-action pipeline: how a system observes the world, builds a representation, and executes physical decisions reliably.

Human-AI Interaction  —  Most AI products reset context between sessions. Mio is an early attempt to build something different — a companion that accumulates history, maintains consistency, and becomes more useful over time rather than starting fresh each conversation.

Autonomous Navigation  —  SLAM, sensor fusion, and motion planning are the technical foundation for anything MECH builds that moves. Working through these systematically as part of the ROS2 and robotics learning track.

Agent Architectures  —  How do you build a system where multiple specialized agents coordinate over shared state without becoming brittle? AetherXD explores this at the desktop level.

Embedded Intelligence  —  Running capable inference on constrained hardware without a cloud fallback. Central to making MECH IoT and hardware products viable as standalone devices.

---

Engineering Roadmap

mermaid
timeline
    title Engineering Roadmap
    section Building Now
        Mio v1           : Fine-tune · local inference · companion layer
        Reception MVP    : Visitor AI · appointment engine · India launch
        AetherXD core    : Agent runtime · desktop shell · model orchestration
    section Learning Now
        Rust             : Systems programming · MECH infrastructure
        ROS2             : Robotics middleware · hardware stack foundation
        Deep Learning    : Formal theory · building from scratch
        Autonomous Sys   : SLAM · sensor fusion · control theory
    section Mid-term
        Doraemon         : Persistent memory · task automation · system control
        Robotics stack   : ROS2 + sensor fusion + control loop integration
        DocaMech v1      : Production Windows app · licensing system
        MECH infra       : Shared APIs · cross-product data contracts
    section Long-term
        Robot Arm        : AI-controlled physical automation
        MECH Core IoT    : Smart home firmware platform
        Autonomous sys   : Closed-loop perception to action systems
        Research output  : Architecture writeups · selective OSS releases

---

GitHub Analytics

<!-- Trophies -->

<!-- Stats + Streak side by side -->

&nbsp;

<!-- Profile details full width -->

<!-- Languages side by side -->

&nbsp;

<!-- Activity graph tokyonight colors -->

<!-- Snake animation — runs from output branch via GitHub Actions (see .github/workflows/snake.yml) -->

---

Collaboration

Open to working with people who are specific about what they're building and why.

| | |
|:--|:--|
| Researchers | Local inference, embodied AI, robotics perception, agent architectures |
| Engineers | Desktop software, embedded systems, AI infrastructure, robotics middleware |
| Founders | Technical co-building and product architecture for AI and robotics startups |
| Hackathon teams | Full-stack AI integration under real constraints and tight deadlines |

Formats: research collaboration &nbsp;·&nbsp; selective co-building &nbsp;·&nbsp; technical advisory &nbsp;·&nbsp; hackathons

<a href="https://github.c

…

Collapse