Project Tamara

Inspiration I built Tamara because every AI demo I watched needed accounts, API keys, and good Wi-Fi — things a 16-year-old student can't always count on. I wanted an assistant that works the second you open it, and grows with you: a study buddy for beginners today, a council of four AI brains when you're ready. Judges shouldn't have to trust a demo — so Tamara shows its failures on screen instead of hiding them.

What it does Tamara is a personal AI assistant with two modes. Out of the box it is an offline study buddy: 60-second explainers, marked tests, flashcards, past-paper hits, video scripts, notes, todos, reminders, and lifetime cost tracking — no account, no key, no internet. Start two free servers and it unlocks advanced mode: four AI brains (local Ollama, OmniRoute gateway, NVIDIA NIM, OpenCode) answer side by side in streaming panes with browser voice, and a Council merge produces one verdict. Every chat is saved in a sidebar, and a permanent counter proves the session costs $0.

How we built it The whole app is a single Python file — no framework, no build step. The terminal UI, an anime.js web interface (streaming panes, chat sidebar, synced browser speech), and a Tk fallback all talk to the same four-brain fan-out core with one shared deadline, so fast brains display first and nobody waits for the slowest. Memory, tasks, and notes live in local SQLite plus an Obsidian vault. A 28-test suite guards every feature, including a pollution guard that fails any test touching real user files.

Challenges we ran into Small local models drown in long context, so prompts are trimmed middle-out and memory recall shrinks with prompt length. Streaming four brains at once without thread leaks took a bounded queue with drop-oldest semantics. Browser speech used to read markdown tables as "dash dash dash," so speech now gets a cleaned copy while the screen keeps the full text. The biggest fight was honesty: every brain gets a loud [BRAIN-ERROR] marker instead of a silent fallback, and the whole demo had to still work when brains die on stage.

Accomplishments that we're proud of Four live brains answering at once with a merged council verdict, on a 6GB RAM machine, for $0. A web UI with per-chat saved history that survives reloads. A voice toggle that kills speech mid-sentence and remembers your choice. Offline study tools that genuinely need zero setup. And a demo video cut entirely with project-built tooling — transcribed narration, burned captions, real app screens.

What we learned Constraints are features: 6GB of RAM forced the one-deadline fan-out, the offline-first design, and the $0 architecture. Tests are a demo tool — the pollution guard and thread-leak tests caught real bugs that would have died on stage. And showing failures beats hiding them: judges trust a labeled error pane more than a suspiciously perfect demo.

What's next for Project Tamara More council brains (any OpenAI-compatible endpoint already plugs in), smarter memory recall over the Obsidian vault, study-plan tracking across sessions, and a packaged installer so a beginner can double-click Tamara with zero setup.

PS GUYS BARE WITH MY VIDEO AS I WAS SICK AND I AM NOT ABLE TO DO MUCH