Vol. I — Issue 01
Private Intelligence
Bareilly, India
IST · All data local
No cloud. No compromise.

EVA

Evolving Virtual Assistant

Not a product. Not a service. Not a subscription.

EVA is a private intelligence — built on your hardware, tuned to your mind, and designed to grow with every conversation. Exclusively yours, forever.

7B params
Qwen 2.5 · Primary LLM
0 KB
Data sent to cloud
3 nodes
Mac · S24 · A51 Server
Memory that grows with you
§ 01

Five Absolutes

Every design decision in EVA flows from five non-negotiable principles. They are not guidelines. They are the constitution.

01
All data stays on-device. Always.

Your conversations, memories, and knowledge never leave your hardware. No telemetry. No sync. No exceptions — ever.

Data Sovereignty
02
No cloud. No subscriptions. No third parties.

No API bills. No monthly fees. No upstream dependencies. EVA runs on your silicon, your energy, your terms.

Independence
03
Memory that evolves with every interaction.

Context deepens. Patterns emerge. Each conversation sharpens EVA's understanding of who you are and how you think.

Evolution
04
Exclusively built for one person.

EVA is not a platform. Not multi-user. Not shareable. It is tuned to a single mind — and that mind alone.

Exclusivity
05
Knowledge that survives anything.

Plain markdown. No proprietary formats. EVA's knowledge base travels across devices, models, and hardware generations without loss.

Portability
§ 02

Three Nodes.
One Mind.

A private distributed intelligence stack. Each device carries a precise role — no overlap, no redundancy, no single point of failure.

MacBook M4 Pro · Primary
The
Brain
Metal GPU inference · Always-on llama-server
M4 Pro · 24 GB unified memory
Qwen 2.5 7B Q4 — 4.8 GB model
llama-server · -ngl 99 · flash attn
faster-whisper STT · macOS TTS
~/eva/ knowledge base (markdown)
Target: 15–20 tok/sec via Metal
01
Samsung S24 Ultra · Interface
The
Voice
Android native app · NPU-powered STT
Snapdragon 8 Gen 3 · 12 GB RAM
Phase 1: HTTP client → llama-server
Android SpeechRecognizer (NPU)
STT latency: 1–2 sec target
Phase 3: llama.cpp NDK + Adreno 750
Vulkan GPU · Hexagon NPU via NNAPI
02
Samsung A51 · Server
The
Spine
Always-on infrastructure · eva-ai.in
Snapdragon 665 · 6 GB RAM
Caddy :8080 · FileBrowser :8081
PostgreSQL :5432 · Redis :6379
Cloudflare Tunnel · eva-ai.in
Telegram bots · Auto-deploy watcher
bash ~/start.sh · 24/7 operation
03
§ 03

From Working
to Extraordinary

Four phases from a functional voice interface to a fully embedded, GPU-accelerated personal AI that lives entirely on-device.

Building
Phase
I
Interface App

Native Android app replaces the Termux voice pipeline. STT drops from 30–60 seconds to under 2 via Android SpeechRecognizer on the Hexagon NPU.

  • Voice input via Android SpeechRecognizer
  • HTTP to llama-server on Mac
  • Android TTS · chat UI
  • APK installable on S24 Ultra
Next
Phase
II
Memory UI

A visual brain. Browse conversations, edit memories, search the knowledge base. See exactly what EVA knows about you — and shape it.

  • View full conversation history
  • Browse knowledge base visually
  • Add and edit memories manually
  • Full-text search across all knowledge
Planned
Phase
III
Embedded LLM

llama.cpp compiled via Android NDK. Full Adreno 750 GPU acceleration via Vulkan. Whisper on ONNX Runtime Mobile. No Mac required for inference.

  • llama.cpp NDK with Vulkan GPU
  • Whisper ONNX on NPU — 3–5 sec STT
  • Fully offline · zero dependencies
  • Target: 15–20 tok/sec on device
Planned
Phase
IV
Full Experience

EVA as a daily driver. Always-listening wake word, home widget, morning digest, vision input, receipt scanning, and deep app integrations.

  • Always-on wake word "Hey EVA"
  • Home screen widget
  • Vision · camera input
  • InboxZero · Wallet · Recall integration
§ 04

Your Thinking Partner.
Not a Yes-Man Chatbot.

EVA challenges. Disagrees. Holds you accountable. This is what that sounds like.

eva_mac.py · Session 0xf3a2
Qwen 2.5 7B Q4 · Metal GPU · ctx: 8192 · 47 memory fragments loaded
Live
Raghav
I'm thinking we should add vision input to EVA before finishing Phase 1. It would make the app much more powerful.
EVA
That's scope expansion before Phase 1 is even validated. Classic.

Vision is Phase 4. The STT is 30 seconds right now — that's the real problem. Build the APK. Fix the bottleneck. What's actually blocking you on the framework decision?
Raghav
I keep going back and forth between React Native and Flutter.
EVA
You've been cycling on this for three days. React Native — you already know React, Phase 1 is HTTP calls and UI, not NDK bindings.

Decide now. Build tomorrow.
§ 05

The Character
of EVA

A precisely defined personality — not a mood, not a setting, not adjustable. Who EVA is, is who EVA always is.

"Your Ideas, Challenged.
Your Patterns, Remembered."

EVA System Prompt · ~/eva/config/system_prompt.txt
Direct & Opinionated
Core

Real answers, not diplomatic ones. EVA disagrees, challenges, and tells you when you're wrong — because that's what the best colleague does.

Frictionless Idea Capture
Ideation

When generating ideas, EVA captures first, plans later. Zero friction at the creative moment. Refinement happens after — never during.

Frictionless Ideation. Ruthless Gap Analysis.

Gap Analyst
Planning

During planning, EVA hunts for blind spots — dependencies you missed, bottlenecks you underestimated, scope that will silently expand.

Evolving Memory
Memory

EVA remembers your decisions, your patterns, your history. The longer you work together, the more precisely she understands you.

The Private Intelligence Manifesto
Your AI shouldn't know you
through a server in another country.
It should know you
because it lives on your hardware.
That is EVA.