everyday-value-agent

A family of small AI agents that live inside one laptop. No cloud, no telemetry, no API keys — every model runs on the machine you're sitting at.

EVA is a long-running experiment in a specific constraint: how good can a personal assistant get when the entire thing has to fit in 4GB of VRAM and 5GB of free RAM? The answer so far is "further than you'd think, if you stop asking one big model to do everything."

Not one model with a long prompt — a society of small ones, wired so the slow parts never block the fast parts:

  you speak ──► speech-to-text (Parakeet int8, CPU, ~245ms)
                        │
              ┌─────────┴──────────┐
              │  deterministic     │   "open github", "how much battery",
              │  fast-paths + kNN  │   "convert 5km to mi"
              │  of your past use  │   ──► answered in ~29ms, no LLM at all
              └─────────┬──────────┘
                        │  (novel phrasing only)
              ┌─────────┴──────────┐
              │  embedding gate:   │   is this a command, or just talking?
              │  chat or command?  │
              └─────────┬──────────┘
                        │
        ┌───────────────┼────────────────┐
        │               │                │
   tiny router     persona chat     deliberate tier
   (qwen 1.5B)     (gemma2:2b)      (phi4-mini + tools)
        │               │                │
      a skill      her real voice    plan → act → observe → self-correct

The tools reason, so the model only has to route. A 3B model asked "if a train leaves at 2:40pm and takes 95 minutes, when does it arrive?" guesses — and gets it wrong. Handed a time_math tool, it returns 4:15 PM exactly. On a checkable question set that's the difference between 75% and 100%.

This is a personal project that's graded honestly rather than marketed. The repos carry their own adversarial graders, and the results — including the bad ones — are committed alongside the code.

• Fast where it matters: ~29ms median on commands, ~245ms speech-to-text.
• Still slow where it isn't: the deliberate tier's worst case is ~21s when four models compete for RAM. That's the current work.
• Diana is not a product. She's wired to one specific machine and says so in her own README. brain is the part built to be reused.

_{Everyday Value Agent · built on a laptop in Sri Lanka 🇱🇰}