Available for work
Initializing…
SAHIL CHANNA
I teach machines to think. Building intelligent systems with deep learning, LLM agents, and data — turning hard problems into shipped software.
About Me
I build LLM agents that stay up
I'm an AI/ML engineer working on agents that hold real tools, keep state between conversations, and have to keep working when nobody is watching. The proof is Zara — a self-hosted AI assistant with 229 skills and five orchestrated sub-agents, running 24/7 on a Raspberry Pi 5 on my desk. I wrote every line of her.
Most of what I know came from her breaking. Inference now fails over across three providers because one of them went down mid-conversation. She scores herself nightly on failure-honesty because she used to report success on work she had never done. I'm an undergraduate studying AI and ML, but the deep learning, NLP, and retrieval work I care about is the kind that ships and then has to survive contact with a real user — me.
Technical Skills
Python
Advanced proficiency in Python for data science and machine learning
TensorFlow
Building and deploying deep learning models with TensorFlow
PyTorch
Developing neural networks using PyTorch framework
Scikit-learn
Machine learning algorithms and model evaluation
Pandas/Numpy
Data manipulation and numerical computing
AWS/GCP
Cloud computing and deployment on AWS and GCP
Docker
Containerization and deployment
More
Continuously expanding my technical toolkit
Selected Work
An index of things I've built — each one a real problem I wanted to solve. Hover a row; click through to the code.
Stream Select JavaScriptEdTechBilingual
A two-round self-rating quiz that helps BTech and MBA students find which stream actually fits them. Zero-framework vanilla JS, English/Hindi toggle, live on its own domain.
ARCL CLI Node.jsCLILLM
Transactional coding CLI that edits files via unified diffs. AI-assisted code mutations with automatic backups, rollback, and multi-provider LLM support.
Mareen PythonAIVoicePrivacy
Privacy-focused voice assistant with immersive 3D orb UI. Offline speech recognition, Hindi support, and local LLM processing with Ollama.
PaperPal PythonAIResearch
AI-powered research paper assistant that helps you discover, analyze, and summarize academic papers efficiently.
Meet Zara
My longest-running build — a personal AI that lives on a Raspberry Pi 5 on my desk. Not a chatbot demo: she runs 24/7 across a dozen services, holds 229 skills, remembers everything, and actually gets things done.
What she is
Zara is a self-hosted AI assistant I've been building from scratch in Python. Her brain runs on a Raspberry Pi 5, with a second presence on my laptop as a glowing desktop orb. She talks over Telegram, voice, and a web UI. Inference sits behind a shim that fails over between three backends: NVIDIA NIM, Ollama's cloud, and a fully local model when everything else dies.
She isn't one app; she's a small ecosystem: a persistent memory layer synced between devices, an orchestration layer that routes any request to the right one of five specialised sub-agents — and lets them hand work to each other mid-task — and an automation stack that briefs me every morning, watches her own health, and improves her own code at night.
Her memory is a MAG stack — Memory-Augmented Generation — layered rather than one store: eight markdown files injected hot into every prompt, FTS5 keyword search, local vector embeddings, structured facts, and episodic day files. Verified on the Pi, 18 Aug 2026: 12,597 turns indexed and 12,616 embedded — 99.94% coverage, not 100%, because embedding runs behind live writes and always trails by a few. Recall fuses the keyword, semantic and fact arms by reciprocal rank instead of adding scores on incompatible scales — which let whichever arm had the widest range quietly win — then re-ranks for recency and diversity, under 40 tests. The embeddings come from a local model and never leave the device.
The hardest part wasn't making her capable — it was making her honest. A grounding guard, a verify gate that demands a real exit code, and a nightly eval that scores her on failure-honesty and sycophancy all exist because she used to report success on work she had never done.
The Zara ecosystem
Telegram Brain
The always-on core, running on the Pi. Chat, Hinglish voice notes transcribed by faster-whisper, reactions to any media you send, and reminders in plain language. Message edits are paced and back off on rate limits, so a throttle costs latency and never the reply.
Desktop Orb
A glowing orb companion on my laptop with its own local LLM on the GPU. Streaming voice in and out — she starts speaking the first sentence while still thinking of the rest.
Zara OS
A React PWA front-end with passkey login — chat, skills, permissions, and a live graph of her memory, all in the browser.
Agent Orchestration
A router decides which of five sub-agents owns a request — Rhea codes, Zoya researches, Sia reasons, Myra recruits, Nova runs ML. They delegate to each other and merge results; one has already found and fixed a bug in another's pipeline, live, without me touching it.
Career Copilot
Send her a job link on Telegram and get back a one-page resume PDF tailored to that posting, with every claim verified against the real data. She scans job boards too — but never applies without me.
Nervous System
Morning briefings, an n8n workflow stack, and a watchdog that tails every service log, classifies the failure, restarts what died and pages me on Telegram. A sentinel hashes her own code and skills every few minutes — when something changes, she notices and tells me at the next turn.
Honesty & Evals
Daily gates that score her on grounding, failure-honesty and sycophancy, plus a coder eval that fails the build on fabrication. A verify gate demands a real exit code before she says "done", and she now retracts a claim when no tool actually performed the action.
Zara Code
A Claude-Code-style coding agent in the terminal — Ink TUI over a Python engine, plan mode, permission gates, live step streaming. Her coder now also drives the real Claude Code CLI, sandboxed in its own git worktree so a runaway edit can't touch main.
Zara Craft
A Three.js voxel survival game where Zara is your in-game companion — endless nights, mob waves, perk milestones, and full touch controls on the phone.
Why not just Alexa, Siri, or ChatGPT?
Fair question. Here's where a self-built assistant beats the big ones — for me.
She runs on a Pi 5 and my laptop — no company server sitting between me and my assistant. Alexa and Siri are microphones wired to someone else's cloud.
Persistent memory synced across every device she lives on. Each conversation builds on the last, instead of resetting per app or per session.
She pushes git commits, builds PDFs, runs automations, and manages her own repos — assistants that only answer questions are half an assistant.
Hinglish-native, voice or text, tuned to how I actually talk — not a one-size-fits-all persona.
Three inference backends behind one shim — NVIDIA NIM, Ollama's cloud, and a local model on the Pi. When one dies she fails over instead of disappearing. An assistant you depend on can't have an off switch you don't own.
A grounding guard and a verify gate stand between "I did it" and actually having done it, and a nightly eval scores her on failure-honesty. An assistant that quietly reports success is worse than one that says no.
A nightly loop reviews her own failures against a baseline and patches what's weak. She's better this month than last month, by her own hand.
And the open-source crowd?
OpenClaw and Hermes are great self-hosted agent frameworks — Zara plays in the same league, but she's hand-built for one human, not installed from a repo.
| Zara | OpenClaw | Hermes | |
|---|---|---|---|
| What it is | A personal AI ecosystem, written from scratch in Python | Open-source agent runtime + message router | Self-improving agent daemon by Nous Research |
| Runs on | A Raspberry Pi 5 + my laptop, 24/7, hardware I own | Your machine or a VPS (Node.js service) | Your server or WSL2, as a daemon |
| Reach | Telegram, Discord, a voice orb, a web PWA, the terminal | ~29 chat channels | 16+ messaging platforms |
| Models | Three backends behind one shim — NVIDIA NIM, Ollama cloud, local Ollama — with bidirectional failover | Bring your own — Claude, GPT, Gemini, or local | Nous models, plus bring-your-own |
| Memory | Persistent + two-way synced, with local embeddings, rank-fused retrieval, and a visual memory graph | Long-term memory on your machine | Persistent memory that grows with use |
| Gets better by | A nightly self-improve loop, a 5-agent team that delegates work, and daily eval gates that fail her build on fabrication | Writing its own new skills | Writing reusable skill docs from solved problems |
| Beyond chat | A coding agent, a voxel game she stars in, a career copilot, home automations | Browsing, files, commands, scheduled tasks | Cron tasks + tool use |
| How you get it | You can't — she's mine. Every line written and tuned for one person | Install the framework, add API keys | One-line install script |
What building her taught me
Not a tutorial project — I learned each of these because something in her broke without it.
Giving a model real hands — and handling the calls that come back wrong.
Routing work across five sub-agents that delegate to each other mid-task.
Plan, act, check, retry — how she heals and improves her own code.
A protocol shim in front of three providers, picked per task, failing over when one dies.
Rank fusion over keyword + embedding + fact arms, and deciding what's worth storing at all.
Quantised models and local embeddings inside a Pi 5's memory budget.
Every claim on a generated resume checked against real data.
Scoring honesty and sycophancy on a schedule — the failure mode you can't see is silent success.
Watchdogs, n8n schedules, passkeys — keeping 24/7 actually 24/7.
Get In Touch
Have a project in mind or want to discuss potential opportunities? Feel free to reach out — I usually reply within a day.
Contact Information
Location
India (Open to remote work)
GitHub
github.com/saxillResume
Download PDFResponse Time
Usually replies within 24 hours