Skip to content
View Sachitt-AV-08's full-sized avatar
9 followers · 3 following

Achievements

Achievement: Pull Shark

Achievements

Achievement: Pull Shark

Highlights

Block or report Sachitt-AV-08

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Sachitt-AV-08/README.md

A V Sachitt

Building local-first AI that refuses to guess.

> whoami
A. V. Sachitt  |  17  |  1st year college
IIITDM Chennai (Smart Manufacturing) + IIT Madras (Management and Data Science)

Website LinkedIn Email Instagram M8ven Visitors

The stack I'm putting most effort into

An agent that can act on your machine should not be able to act on anything. These three are built so that each one can stop the next.

  goal ──▶ errands ──── is there a checked recipe for this?
                │              no ─▶ refuse
                │              yes
                ▼
          sentinel ─────── is this action reversible?
                │              no ─▶ ask a human
                │              yes
                ▼
           orvima ──────── act on your real browser,
                            in a viewport you can pause
orvima MCP server. Any agent gets browse_* tools on your own Chrome, Edge or Brave - your logins, your machine, your live viewport. 22 tools, 470 tests, MIT.
sentinel Calibrated local classifier. Decides which actions run unattended, which need a human, and which are refused. Fails closed, including when it is broken.
errands Recipe-based goals. A goal matches a checked recipe or it is refused. Every step gated, every step verified. No generative planning.

Featured work

graphify

74,084 lines  |  Python

Codebase knowledge graph. Tree-sitter extraction across 36+ languages, an MCP server, and integrations for 20+ AI assistants. The largest thing here, by a wide margin.

MIT · source

karna

56,113 lines  |  Python

Neural core with trained weights and a BPE tokenizer. 229 Python modules. Private - ask for access.

private · ask for access

orvima

14,504 lines  |  Python

MCP server giving any AI agent eyes and hands on your own Chrome, Edge or Brave. 22 tools, a live viewport you can pause, and 470 tests.

MIT · source

AUDIT

13,247 lines  |  TypeScript

One shell, many worlds. TypeScript toolkit for running autonomous digital-intelligence environments.

MIT · source

vijay-voice-assistant

6,310 lines  |  Python

Modular desktop voice assistant. 30+ capabilities with wake-word detection, running fully local.

MIT · source

parley

4,436 lines  |  Python

Drives your already-installed WhatsApp Desktop locally. No cloud, no QR, no second browser profile.

MIT · source

rotation-scan-3d

3,928 lines  |  Python

3D body scanning from a phone rotation video. Pure geometry, zero neural networks, runs on a laptop CPU.

MIT · source

coda-forge-3d

3,070 lines  |  Python

18-stage photorealistic human reconstruction from phone video. COLMAP, Gaussian splatting, geometric fallback.

MIT · source

bhaasm-transformer

2,535 lines  |  Python

From-scratch decoder-only transformer. RoPE, grouped-query attention, SwiGLU, optional MoE.

MIT · source

deception-engine

2,076 lines  |  Python

Defensive honeypots and canary tokens for adversary misdirection. MITRE ATT&CK mapped.

MIT · source

coda-security-primitives

1,866 lines  |  Python

AES-256-GCM, Argon2, PBKDF2, MFA and tamper-evident audit chains. No dependencies.

MIT · source

genesis-media-engine

1,714 lines  |  Python

Declarative DSL for rendering video, audio and images through FFmpeg.

MIT · source

claro-studio

14,718 lines  |  TypeScript

AI-native IDE for a small reversible UI language, with live visual editing. Private.

private · ask for access

The rest of the 44 repositories

Large systems

Project Lines What it does
graphify 74,084 lines Codebase knowledge graph tool ΓÇö 36+ languages, 9+ LLM backends, MCP server, 20+ AI assistant integrations.
priva 23,150 lines .
orvima 14,504 lines Why should Claude be the only one with a browser? Orvima gives any MCP-capable AI (Claude, Cursor, Copilot, your own agent) eyes and hands on your own Chrome or Edge ΓÇö your logins, your machine, your watch.
AUDIT 13,247 lines AUDIT ΓÇö Autonomous Universal Digital Intelligence Toolkit.
vijay-voice-assistant 6,310 lines Modular desktop voice assistant with 30+ capabilities, wake word detection, and real-time processing.
parley 4,436 lines Drive your already-installed WhatsApp Desktop locally: read, send, reply, react & schedule ΓÇö no cloud, no QR, no new browser tab.
rotation-scan-3d 3,928 lines No-GPU 3D scanning from phone rotation videos ΓÇö pure geometry, no neural networks, no cloud.
errands 3,790 lines Recipe-based goals: a goal matches a checked recipe or it is refused.

COSMIC ecosystem - composable pieces, each usable on its own

Project Purpose
Cosmic-Core The shell the other pieces plug into
Cosmic-Explore Graph-powered file explorer
Cosmic-Atlas 3D codebase knowledge graph
Cosmic-Debug Visual debugger, traces through time
Cosmic-Flow Drag-and-drop pipeline builder
Cosmic-Voice Local speech in, speech out
Cosmic-Vision Gesture recognition and hand tracking
Cosmic-Canvas Sketch to 3D
Cosmic-Shield Supply chain security for your deps
Cosmic-Connect Local service registry

Small tools - the ones that get used daily

Project What it does
servehere File server with a QR code  ·  113 lines
patchwork One CLI to set, get and diff YAML/JSON/TOML  ·  264 lines
envcheck Scans code and .env for missing keys  ·  182 lines
pixelflow Terminal image toolbox  ·  204 lines
picgrid Lays out N images in a labelled grid  ·  docs & design
imagediff Before/after slider for model output  ·  docs & design
modelview Drag-and-drop GLB/OBJ/STL viewer  ·  docs & design
neural One CLI for local and cloud LLMs  ·  243 lines
All 40 public repositories
Repository Lines Description
graphify 74,084 lines Codebase knowledge graph tool ΓÇö 36+ languages, 9+ LLM backends, MCP server, 20+ AI assistant integrations
priva 23,150 lines
orvima 14,504 lines Why should Claude be the only one with a browser? Orvima gives any MCP-capable AI (Claude, Cursor, Copilot, your own agent) eyes and hands on your own Chrome or Edge ΓÇö your logins, your machine, your watch. Self-hosted, local-first, MIT. MCP-native browse_* tools + live viewport.
AUDIT 13,247 lines AUDIT ΓÇö Autonomous Universal Digital Intelligence Toolkit. One Shell, Many Worlds. File import, editing, analysis, TODOs, key phrases, 19 integrated tools. Electron + React + TypeScript.
vijay-voice-assistant 6,310 lines Modular desktop voice assistant with 30+ capabilities, wake word detection, and real-time processing
parley 4,436 lines Drive your already-installed WhatsApp Desktop locally: read, send, reply, react & schedule ΓÇö no cloud, no QR, no new browser tab. Local-first CDP automation for Windows/macOS/Linux.
rotation-scan-3d 3,928 lines No-GPU 3D scanning from phone rotation videos ΓÇö pure geometry, no neural networks, no cloud
errands 3,790 lines Recipe-based goals: a goal matches a checked recipe or it is refused. Every step gated and verified, no generative planning.
agent-bootstrap 3,586 lines One command turns any machine into a 10/10 AI agent environment: 163 role-scoped agents, fleet orchestration, plan-to-PR pipeline. Free-tier, local-first.
sentinel 3,432 lines Calibrated local action classifier: decides which browser actions run unattended, which need a human, and which are refused.
coda-forge-3d 3,070 lines 18-stage photorealistic 3D human reconstruction pipeline with CPU geometric reconstruction and Gaussian splatting
Cosmic-Core 2,855 lines Local AI operating system ΓÇö the life force of your COSMIC device
bhaasm-transformer 2,535 lines From-scratch decoder-only transformer with RoPE, GQA, SwiGLU, and optional MoE ΓÇö 2.2B parameters
Cosmic-Vision 2,411 lines Gesture recognition engine ΓÇö sees what you mean before you touch
coda-financial-agent 2,385 lines AGNI/VEDA trading agent with paper trading, backtesting, LSTM prediction, and sector analysis
apkscope 2,152 lines MCP server to view and drive a running Android app as text, for agents without image input
deception-engine 2,076 lines Defensive honeypot/deception framework for adversary misdirection and threat intelligence
Cosmic-Voice 1,881 lines Voice pipeline ΓÇö local-first speech, no cloud, no API keys, just your voice
coda-security-primitives 1,866 lines Production-ready encryption, password hashing, and key derivation for Python applications
genesis-media-engine 1,714 lines DSL-driven media rendering engine for video, audio, and image generation
Cosmic-Debug 1,321 lines Visual code debugger ΓÇö traces through time, finds the bug
Cosmic-Explore 1,266 lines Visual Architecture Yielding Understanding ΓÇö graph-powered file explorer that flies through code
Cosmic-Shield 1,145 lines Supply chain security platform ΓÇö the armor for your dependencies
Cosmic-Connect 982 lines Local service registry ΓÇö the energy center of your COSMIC device
Cosmic-Atlas 928 lines Interactive 3D codebase knowledge graph ΓÇö navigate code like star constellations
Cosmic-Flow 892 lines Visual pipeline builder ΓÇö threads of execution, drag-and-drop
Cosmic-Canvas 669 lines AI sketch-to-3D ΓÇö gives form to your ideas
nebula-fs 554 lines Force-directed file explorer -- navigate your filesystem as nodes and edges, not folders and lists
patchwork 264 lines Config file patcher ΓÇö set/get/diff across YAML, JSON, TOML from one CLI
neural 243 lines Unified CLI for local + cloud LLMs ΓÇö BHAASM, Ollama, NIM in one command
pixelflow 204 lines Terminal image toolbox ΓÇö ASCII preview, frame extraction, thumbnails, format conversion
envcheck 182 lines Environment variable validator ΓÇö scan code + .env, report missing keys before runtime
servehere 113 lines Instant file server with QR code ΓÇö share files over LAN in one command
.github docs & design Profile README
COSMIC docs & design Connected Operating System for Multimodal Intelligent Computing ΓÇö the COSMIC ecosystem
imagediff docs & design Visual before/after image comparison ΓÇö slider overlay for model output evaluation
JARVIS-by-A.V.SACHITT docs & design
modelview docs & design Drag-and-drop 3D model viewer ΓÇö GLB, OBJ, STL in the browser with Three.js
picgrid docs & design Image grid comparator ΓÇö lay out N images in a labeled grid for A/B comparison
Sachitt-AV-08 docs & design GitHub Profile README

By the numbers

Every figure below was measured by cloning each repository and counting the source. Vendored, generated and minified trees were excluded.

  253,229     lines of code   (all 44 repos, public and private)
  182,175     lines in public repositories
  71,054     lines in private repositories
       44      repositories
       40      public
        4      private
        1      developer
Language Share
Python 71.3%
TypeScript 14.6%
HTML 11.3%
CSS 1.3%
PowerShell 1.2%
Shell 0.2%

Python

TypeScript

HTML

CSS

PowerShell

Shell

JavaScript

Principles

  1. Local-first. Your hardware, your data, your logins. No key, no subscription, no telemetry.
  2. Refuse rather than guess. Unknown is treated as irreversible.
  3. Count, don't claim. Every number in this file was measured from the source.
  4. Verify the layer that reads. A passing suite proves nothing; a mutated one that should fail proves something.

Beyond the code

Table tennis, cricket, football, badminton, chess. City- and district-level rankings in Free Fire and Rocket League. I read, write and speak Sanskrit - seven years of it - alongside English, Hindi, Telugu, Tamil, Kannada, Bengali and Spanish.

Building an AI ecosystem and playing competitive esports turn out to need the same two things: staying calm under pressure, and thinking ten moves ahead.

Current focus

> cat /etc/coda/current
+-- orvima: real-browser agent safety, shipped to PyPI
+-- AGNI/VEDA dual-agent training
+-- IEEE paper: Rotation-Scan + AGNI/VEDA
+-- 3D dataset training (Colab T4 + Kaggle P100)

Built by A V Sachitt  ·  Website  ·  LinkedIn

If something here is useful to you, an issue or a star genuinely helps. If you want to fund the work, open an issue and say so - sponsorships are being set up.

Pinned Loading

  1. bhaasm-transformer bhaasm-transformer Public

    From-scratch decoder-only transformer with RoPE, GQA, SwiGLU, and optional MoE — 2.2B parameters

    Python 3

  2. coda-forge-3d coda-forge-3d Public

    18-stage photorealistic 3D human reconstruction pipeline with CPU geometric reconstruction and Gaussian splatting

    Python 3

  3. coda-security-primitives coda-security-primitives Public

    Production-ready encryption, password hashing, and key derivation for Python applications

    Python 3

  4. deception-engine deception-engine Public

    Defensive honeypot/deception framework for adversary misdirection and threat intelligence

    Python 3

  5. genesis-media-engine genesis-media-engine Public

    DSL-driven media rendering engine for video, audio, and image generation

    Python 3

  6. rotation-scan-3d rotation-scan-3d Public

    No-GPU 3D scanning from phone rotation videos — pure geometry, no neural networks, no cloud

    Python 3