TOMORA

personal future intelligence: explore the futures your choices could create, talk with a future-self persona grounded in your real circumstances,

Comment

Inspiration

I'm a self-taught developer from Hargeisa. I freelance on Upwork, learn AI in the evenings, and my family depends on my income. Every big choice I face (a bootcamp, a degree, staying freelance) feels like a guess about a future I can't see.

AI assistants answer the question you just asked. Nobody helps with the question underneath: who could I become, and which road gets me there with the life I actually have? So I built tomora-ai: a way to meet your future self, the successful you, before you get there.

What it does

tomora-ai is not a chatbot. It's an agentic system: a team of 17 AI agents on NVIDIA Nemotron that remember you, call tools, check each other's work, and keep working 24/7.

It follows one journey: understand you → explore futures → compare and decide → act → adapt.

  • 🧭 Understands you. A short onboarding: your ambition and horizon year, where you are now, what you have to work with, what stands in the way, your values and how you want to be coached.
  • 🔗 Knows the real you. Import your LinkedIn, X and Facebook, and link your email and calendar, so your future self learns how you write and what's coming up. You approve every import and can delete it any time.
  • 🛣️ Future Simulator. Nemotron maps 6 to 12 clearly different roads to your goal, each with time and cost ranges, assumptions and risks. Change your hours or budget and every road recalculates instantly, with real arithmetic.
  • ⚖️ Decisions you can trust. The AI never grades its own ideas. Nemotron proposes the roads, plain-code rules protect your money, hours, dependants and risk, and Jev by TypeSafe AI makes narrow, typed judgments with confidences, in 0.37 seconds for about $0.00008 per decision.
  • 💬 Future Self chat. Watch it think: Nemotron's reasoning streams live while it searches your memory, cites the notes it used, and proposes plans that wait for your approval. Attach a photo or a text file and it reads it.
  • 🏛️ Council and missions. Four future selves debate a hard decision in parallel; missions split a big goal across helper agents with a supervisor reviewing every result.
  • ✉️ Future Reflection. A letter from the person you could become, laid out as a magazine issue. A possibility, never a prediction.
  • 🎙️ In your own voice. Your voice is cloned with ElevenLabs only after you say a consent phrase with three random code words out loud. Then voice mode lets you talk to your future self, and it answers in your voice, with every word written by Nemotron.
  • 🌙 Always on. A night shift writes your morning brief and TOMORA Daily every day, even while you sleep, and scheduled calls ring you.
  • 🔐 You stay in control. Memory is plain Markdown notes you own; every change waits on an approval card; every tool call is logged; every visitor gets a private workspace.

How I built it

  • NVIDIA Nemotron-3.5-Super-VL on Nebius Token Factory powers every agent, using what the model offers beyond chat: reasoning on or off per agent, the live reasoning trace streamed to the screen, JSON-schema structured outputs, parallel tool calling, long context and vision.
  • 17 agents in one Docker Agent file, working through an MCP server that exposes tomora-ai's tools: memory search, web research, plans, delegation, spawning helpers, and proposals for calls and emails. Each agent only gets the tools its job needs; a supervisor and a fact checker review results.
  • Jev by TypeSafe AI is the decision layer under Nemotron, with deterministic rules in front of it.
  • ElevenLabs for voice cloning, speech-to-text and speech. On live calls, ElevenLabs uses my app as a custom LLM, so every word is written by my agent on Nemotron.
  • You.com for web research from the cloud.
  • Next.js 16 and React 19, with SQLite and an Obsidian-compatible Markdown vault for memory.
  • Always on in the cloud with Docker Compose, Caddy and free HTTPS, as a service that restarts after crashes and reboots. Multi-user mode gives every visitor an isolated workspace with daily spending limits.
  • About 300 unit and workflow tests and Playwright browser tests.

Challenges I ran into

  • ElevenLabs won't run a call agent with an instant voice clone and a custom LLM. I found this in live testing, so I built voice mode: speech-to-text, then Nemotron with reasoning on, then speech in your clone, turn by turn.
  • A crash only newer Chrome triggered. A scroll call started returning a Promise, React treated it as a cleanup function, and voice mode crashed right after each reply. I reproduced it, fixed it, and added error boundaries so a failure can never take the page down.
  • Making AI advice safe for real money and real families. The model proposes, but code enforces the hard limits, and Jev's confidences decide when you should take a second look.
  • Long structured outputs sometimes came back short, so the simulator validates every road and refuses partial results.
  • Running a public demo on a small server without a surprise bill: per-workspace and per-server limits on tokens, calls and voice replies.

Accomplishments that I'm proud of

  • It's live and always on: anyone can try it in two clicks, in their own private workspace.
  • Talking to your future self in your own voice, with consent you give out loud.
  • On my 19-scenario decision test, the system never picked a road that breaks your limits, and was correct on all 57 runs.
  • A solo build that feels like a finished product: onboarding, memory, agents, decisions, voice, and a brand of its own.

What I learned

Agents become trustworthy when the boundaries are code, not prompts: tools scoped per agent, approvals before anything changes, rules before judgment, and logs of everything. Use reasoning only where it pays off. And testing against the live services finds problems that mocks never will.

What's next for tomora-ai

  • Shareable clips of your future self's calls
  • A phone call with the whole council of future selves
  • Two users' twins meeting and planning together, with both owners approving
  • A fully private mode on Nemotron Nano, running locally
  • Support for more languages, starting with Somali

Built With

  • agentic-workflows
  • ai-agents
  • dockeragent
  • llm
  • nebius-token-factory
  • nvidia-nemotron
  • persistent-memory
  • personal-ai
  • python
  • tool-calling
Share this project:

Updates

Submission history