From air-gapped AI platforms and model-training pipelines to privacy-minded products and research that has to survive its own experiments. Second-year B.Tech CSE at Siddhartha Academy of Higher Education, Vijayawada.
Public repositories link out; private ones are labelled. Statuses are literal: Specs, Blueprint and Design mean a written design, not shipped code.
AWAN
Active
Private · Halcyon Technologies
A self-hosted, air-gapped sovereign AI platform for confidential industrial and government work. Four strictly separated layers cover model inference, knowledge, agent execution and user interfaces. Submitted to Smart India Hackathon 2026 as team Akatsuki.
Air-gappedOpen-weight LLMsAgentic AISovereign
Not public
Awan is built for places where data must never leave the building. Everything runs on your own hardware with no outbound calls, split into four layers: model inference, knowledge, agent execution and user interfaces. On top sit four products: a chat assistant, a coding assistant, a bot builder and an API. Each layer talks to the next only through written contracts, which is what makes it auditable. It was submitted to Smart India Hackathon.
NEXUS
Active
Public · Earlier version public
A personal AI system on two parallel tracks: a distillation pipeline toward a deployable mixture-of-experts model served on CPU, and a from-scratch research build of the HOPE architecture on a Qwen3-4B base. The first manager-worker version is public; adapters and datasets are private.
NEXUS runs a small manager model that plans work and hands pieces to even smaller worker models, executing the plan as a graph of dependent steps. A skill engine loops until a task passes its own checks, and a vector store gives it long-term memory. The second track is a from-scratch build of a new memory-first architecture on a small base model. The point of the whole thing is to get useful assistant behaviour out of hardware without a GPU.
Language-free thought
Research
Paper in progress
Do populations of tiny language-free learners learn what a single learner cannot? A simulated world of reward-modulated Hebbian entities (about 10k weights each, no backprop) tests the claim, with experiments run on Kaggle and a paper in progress.
The experiment asks whether many tiny learners, each with only about ten thousand weights and no backpropagation, can learn something a single learner cannot. Entities live in a simulated world and adapt by reward-modulated Hebbian rules, the kind of local learning that biological neurons are thought to use. Runs happen on Kaggle. It is a research question, not a product.
Dawn
Specs & contracts
Private · Halcyon Technologies
The container and networking substrate under Awan, specified as five Go components: engine, certificate authority, port registry, link protocol and a single ingress, with written contracts between them.
GoProtobufNetworkingIsolation
Not public
Dawn is the layer that actually runs things: a container engine, its own certificate authority so services trust each other, a registry that hands out ports, a link protocol for service-to-service traffic and a single ingress that is the only door in. The five Go components are specified through written contracts before code, so each can be replaced independently. It exists so Awan can run without Docker.
Zero
Open source
Public · Fork of prime-agent
A coding and research agent built around a Recursive Language Model: a persistent Python REPL, built-in subagents, background sessions and a harness that refines itself from evidence. My rebranded fork of PrimeIntellect's prime-agent.
Zero is a coding and research agent whose main loop is a persistent Python REPL: it writes code to read, search and transform its own context instead of stuffing everything into a prompt. It can spawn sub-agents, keep background sessions running, and refine its own harness from evidence of what worked. It is a fork of an upstream open-source agent, renamed and extended.
Kernel-C training data
Datasets live
Public · Hugging Face
Training data for a kernel-level C coding model. Two public datasets cover kernel and MCU/RTOS source plus systems-programming knowledge, ahead of a pilot fine-tune on Qwen2.5-Coder-14B.
Two public datasets on Hugging Face aimed at teaching a model to write kernel-level C: one holds kernel and microcontroller/RTOS source, the other systems-programming knowledge. They are the input to a planned pilot fine-tune of a 14B coding model. Public, so anyone can train on them.
Solomon
Research & design
Design documents
A formally verifiable, experiment-backed data-analysis engine: every claim must trace to a checkable experiment. Thirty features across thirteen containers, with causal-discovery components prototyped in C and a closed-loop search that improved the Bayesian posterior method.
Causal inferenceVerificationC
Solomon is a data-analysis engine that refuses to state a conclusion it cannot back with an experiment. Thirty features are organised into thirteen containers, with causal-discovery pieces prototyped in C and a closed loop that proposes, runs and checks hypotheses. The interesting part is the discipline: every claim carries a trail to the experiment behind it.
Armo
MVP architecture
Private · Halcyon Technologies
A security-first, AI-assisted operating system for Indian advocates and small legal teams. Envelope-encrypted documents, Indian-script OCR and grounded legal retrieval, designed so a busy advocate finds it immediately practical.
FastAPINext.jsPostgreSQLCloud KMS
Not public
Armo is an operating system for advocates and small law offices: documents are encrypted per file, Indian-script scans are read by OCR, and legal questions are answered from retrieved sources rather than from memory. The design target is the clerical work that eats an advocate's day. It is a Halcyon team project, currently at MVP architecture.
Omega
Archived
Private · Denvire Tech
A campus-digitalisation platform that puts student, teacher, admin and parent workflows on one backend: a locked-down Avalonia desktop app for students, React web apps for staff and parents, an ASP.NET Core API on PostgreSQL, Python services for plagiarism checks and autograding, with OpenFGA for authorisation and Judge0 for code execution. Co-developed with a teammate, then split into ten repositories in July 2026; the organisation is now archived.
.NET 10ReactPostgreSQLOpenFGAFastAPI
Not public
Omega puts a whole campus on one backend. Students get a locked-down desktop app built with Avalonia, teachers and admins use React web apps, parents get a lighter web view, and everything sits on an ASP.NET Core API over PostgreSQL with Python services alongside. The locked-down student client is the standout piece. It lives in the Denvire Tech organisation, is currently archived.
Trobo Health
Design phase
Concept & design
An AI intake assistant and clinic-management platform for small Indian clinics. The intake concept was submitted to Samsung Solve for Tomorrow 2026; the clinic system is designed around ABDM and DPDP compliance.
Next.jsASP.NET CoreGeminiABDM
Trobo is a clinic assistant for small Indian clinics: patients describe symptoms to an AI intake step before they see the doctor, and the clinic side handles records and scheduling. The design is shaped around India's health-data and privacy rules from the start. The intake concept was submitted to Samsung Solve for Tomorrow 2026. It is private and still in design.
HealthResQ
Collaboration
Public · Co-built, tejo123-HUB
A federated health-resource and supply-chain resilience platform. The operations and facility-systems direction is implemented: inventory ledger, hospital encounters, geography hierarchy and role-scoped terminals.
HealthResQ is about keeping hospitals supplied when something goes wrong. The built part tracks inventory as a ledger, records hospital encounters, models a geography hierarchy and gives each role a scoped terminal. The federated, multi-facility side is the direction it is growing toward. Built together with collaborators.
Tapcard
Built
Android app
An Android business-card manager: a 3D card flip, camera OCR, AI field extraction, and instant sharing over NFC, QR and Nearby Connections. All eight planned build phases are complete.
KotlinJetpack ComposeNFCML Kit
Tapcard turns a paper business card into a contact in seconds: point the camera, on-device OCR reads it, and an AI step (your own key) fills out the fields. Cards flip in 3D and can be shared by NFC tap, QR code or Nearby Connections. It is written in Kotlin with Compose, and all eight planned build phases are complete.
Kart
Prototype
Public repository
A campus commerce prototype spanning food, documents, tasks, student resale and rentals, events, merchandise and services, with simulated KC credits. Includes a role-based Next.js app and a FastAPI backend demo.
Kart is a campus marketplace prototype where one app covers food, printouts and documents, tasks, student resale and rentals, events, merchandise and services, paid in simulated credits. There are different views for different roles, a Next.js front end and a FastAPI backend demo. It is a prototype, and the credits are not real money.
Graviton
Architecture & scaffold
Public repository
A multi-platform finance, budget and asset manager that uses classical ML (forecasting, anomaly detection, Monte Carlo goals) rather than an LLM. Monorepo with a Next.js web app, a Tauri desktop shell and an Axum API.
Graviton manages money, budgets and assets across web and desktop, and deliberately uses classical machine learning instead of a language model: forecasting, anomaly detection and Monte Carlo simulation for goals. It is a monorepo with a Next.js web app, a Tauri desktop shell and a Rust Axum API. The architecture and scaffold are in place; it is not a finished product.
Colloseum
Prototype
Public repository
A frontend for an AI adversarial-reasoning arena: agent fighters attack and defend each other's claims until the strongest argument survives. Live combat log, logic graph and archive, driven by a client-side mock engine.
Colloseum stages arguments as combat. Agent fighters attack and defend each other's claims in rounds while a live combat log, a logic graph and an archive show what survives. It currently runs on a client-side mock engine, so it is a front-end prototype of the idea, not a connected system.
Xtron
Early backend
Private · Personal repository
A Rust backend for a password-vault service: SRP login, WebAuthn passkeys, TOTP, vaults, items and files, and breach monitoring, with migrations and deployment configs in place.
RustAxumPostgreSQLWebAuthn
Not public
Xtron is the backend for a password vault, written in Rust. Login uses SRP so the password never crosses the wire, with passkeys, one-time codes, vaults, items, files and breach monitoring on top. Migrations and deployment configs are written; it is an early backend, not a released service.
Halo
Blueprint
Concept & design
A multi-cloud SaaS concept that automates non-development engineering work (diagnosis, FinOps, security, monitoring, QA) and hands code changes to engineers. The blueprint stands at 104 features across 15 components.
Multi-cloudAgentsFinOps
Halo is a concept for software that handles the engineering work that is not building features: diagnosing incidents, cloud cost control, security checks, monitoring and QA, handing anything that needs a code change to a human engineer. The blueprint is written down as 104 features across 15 components. There is no running product yet.
W
Specification
Design documents
A specification for a Linux-based, Ubuntu-compatible desktop OS: an OS-level broker between AI agents and the system, a credential vault, an OOXML-native office suite and Windows-application compatibility. Seventeen-plus documents.
LinuxRustWineVFIO
W is a specification, not an installable system: a Linux desktop that is Ubuntu-compatible, with a broker between AI agents and the operating system, a credential vault, an office suite built on the OOXML formats and compatibility for Windows applications. Seventeen or more documents describe it.
devtools
Working
Private · Halcyon Technologies
Internal tooling for Halcyon systems. epb reads real frontend and backend code to catch missing, unused and drifting endpoints; a task board coordinates work. Released under the PolyForm Noncommercial licence.
Node.jsDeveloper toolsCI
Not public
devtools holds two small command-line tools with no dependencies. epb reads the real frontend and backend code and reports endpoints that do not match, are unused, or are requested but never built, and computes what a change would affect. tkb is the shared task and status board that agents and people work from. Both run on plain Node.
prompt-complexity-analyzer
On PyPI
Public repository
A local ML utility that scores prompt complexity from 1 to 10 and recommends a model tier without any API calls. Includes a CLI plus dataset-generation and training scripts.
Give it a prompt and it returns a complexity score from 1 to 10 and a suggested model tier, entirely on your machine with no API calls. It uses sentence embeddings plus hand-built keyword features feeding a gradient-boosted regressor, and ships a CLI with the scripts to generate data and retrain. Published on PyPI.
Doc_Classifier
Completed
Public repository
An offline file organiser that classifies and sorts documents, images and media with local models, so nothing leaves the machine.
An offline organiser for messy folders: it classifies documents, images and media using local models and sorts them, so nothing is uploaded anywhere. Finished and self-contained.
Claude Skills
Open source
Public repository
Mini-documentation on how Claude Agent Skills work, alongside the 47+ custom skills I use for research, architecture documents, frontend design and verification.
A public collection of the skills I give Claude: short documentation of how Agent Skills work, plus the custom ones for research, architecture documents, frontend design and verification. Useful as examples of how to structure a skill.
ThinkStrike
Archived
Public repository
A modular learning platform for AI-powered quiz generation, student feedback, teacher dashboards and admin management.
A learning platform with four sides: AI-generated quizzes, feedback for students, dashboards for teachers and management for admins, built as separate modules. Archived.
02
Teams I build with
Two GitHub organisations, one active and one archived.
Halcyon Technologies
Member · three-person team · active
The home of Awan, the air-gapped AI platform, and Dawn, the container and networking substrate beneath it, plus the legal-OS Armo and the internal devtools that keep the stack honest.
Home of Omega, a multi-tenant campus-digitalisation SaaS for Indian colleges, with portals for students, teachers, admins and parents. Eleven repositories spanning an ASP.NET Core backend, React web apps, an Avalonia desktop client and AI services. Since superseded and archived.
Omegacampus-backendcampus-admin-webcampus-student-desktopcampus-ai-services+ 6 more
03
Build log, 2026
Reconstructed from repository activity, Hugging Face publish dates and project notes. Months are when the work was visibly under way, not launch dates.
B.Tech begins
Computer science and engineering at Siddhartha Academy of Higher Education, Vijayawada, class of 2029. Everything below is built alongside coursework.
NEXUS begins
A local, sovereign assistant that understands Telugu, English and Tenglish, designed for a no-GPU laptop.
Three builds on paper
Tapcard, Trobo Health and the Aegis firewall each get a full plan; the project inventory is written down.
NEXUS goes multi-agent
Manager-worker system, a LoRA fine-tune and an SFT dataset pipeline for a rented L4 GPU.
Open-source push
prompt-complexity-analyzer reaches PyPI, Claude Skills and Doc_Classifier go public, and this site is redesigned.
Graviton and Denvire Tech
The finance-manager monorepo appears, and a campus-platform team spins up under Denvire Tech.
Campus platform sprint
Eleven repositories across backend, web, desktop, parent portal and AI services, plus HealthResQ and Colloseum.
Halcyon and SIH
Armo and the Halcyon repositories begin; AWAN is entered into Smart India Hackathon 2026 as team Akatsuki.
Halo and W on paper
Halo, a multi-cloud engineering-automation blueprint, and W, a specification for an Ubuntu-compatible desktop OS, are written down as documents before any code.
Agents and kernels
Kart and Zero land, Solomon's causal research runs, and the kernel-C datasets are published to Hugging Face.
Gemma-based model scoped
An enhanced model on a Gemma 4 12B base is planned for data analysis, code and shell execution with reasoning on the results, and metacognition. Scoped, not shipped.
Awan, Dawn, Prani
Halcyon's Awan, Dawn and devtools are under active commits, the language-free thought paper takes shape, and weight-editing experiments begin.
This site, in numbers
782 commits and 156 merged pull requests across default branches, with roughly 250K lines of code in the repositories I build in, counted 13 Nov 2026.
04
Research notebook
Threads that are not products yet: experiments, specifications and papers. Each is at the design or experiment stage unless the card above says otherwise.
~/lab — bash
ruthvik@lambda1:~$ ls ~/lab
prani/ language-free population learning, paper plus Kaggle runs
solomon/ causal discovery: Order-MCMC with sBIC, tetrad proxy tests in C
Systems architectureAPI and protocol designThreat modellingCausal inferenceClaude Skills authoringTechnical specification
Siddhartha Academy of Higher EducationB.Tech in Computer Science and Engineering, Vijayawada
Graviity Junior College
Westberry Tenali
06
At a glance
The short, factual version: who I am, what I build and how to reach me.
Who
Ruthvik Anne (also written Anne Ruthvik), software developer and a B.Tech CSE student (2025–2029) at Siddhartha Academy of Higher Education, Vijayawada, India.
What I build
AI platforms and agent systems, model-training and data pipelines, full-stack web apps, Android apps and security-first backends.
AWAN (air-gapped sovereign AI), NEXUS (local multi-agent assistant), Armo (AI operating system for advocates), Omega (campus digitalisation), Tapcard (Android business-card app), prompt-complexity-analyzer (on PyPI).
Open to
Internships, research collaborations and product work: web and Android apps, AI and LLM integration, and backend services. Get in touch by email below.