akash@portfolio ~
› establishing secure channel...
AB
Akash Bishnoi
HomeAboutSkillsExperienceProjectsServicesFit CheckCertificatesBlogToolkitContact
All Projects
AI · Case Study

Nexus — AI Personal Assistant

Multi-modal AI with private memory and tool use.

A multi-modal assistant with voice, vision, and a tool-use planner. Context-aware, with private memory vaults under user control.

LLMRAGVoiceVector DB

Product Demo

Interactive workspace

Explore the product workflow with a fixed local dataset. Controls and state changes are functional; the demo makes no external requests and requires no account or API key.

NexusResearch workspace
4 sources indexed
AB
Grounded in selected sources

Grounded answer

Question: What drives retention in enterprise accounts?

Evidence checked

Early adoption and operational fit are the strongest retention signals

Accounts that complete a structured onboarding and adopt three core workflows in the first month retain at materially higher rates. Interview evidence suggests operational ownership is more predictive than team size.

Key findings
  1. 1Accounts adopting three workflows by day 30 retained at 91% after twelve months.
  2. 2A named operations owner reduced time-to-value by a median of nine days.
  3. 3Guided onboarding had the strongest effect in accounts above 250 seats.

Source scope

4 of 4 included

The Problem

General-purpose chat assistants forget you between sessions and ship your data to whichever provider answered last. Power users want continuity, privacy, and real action.

The Solution

Nexus combines a long-term memory vault, a planner that orchestrates tools, and a privacy-first model router. Every memory is local-first and encrypted; the user owns the keys.

Key Features

Private Memory Vault

Encrypted vector store keyed to the user's device — long-term recall without vendor lock-in.

Tool-Use Planner

Decomposes goals into tool invocations with verification steps before execution.

Voice & Vision

Whisper-quality STT, low-latency TTS, and on-device vision for screenshots and documents.

Model Router

Routes each task to the best model by latency, cost, and privacy tier — automatically.

Workflows

Reusable, shareable agent flows that chain tools, prompts, and human checkpoints.

Provenance Trail

Every output links back to the exact memory chunks and tools used to produce it.

Architecture

  1. 1Local-first vault using SQLite + sqlite-vss for offline RAG; sync over CRDT to encrypted cloud snapshots.
  2. 2Planner emits a typed action graph; executor validates each tool call against a JSON-schema allowlist.
  3. 3Router scores models on a rolling EMA of latency / quality / cost per task class.
  4. 4Provenance store anchors every response token to its source chunks for full traceability.

Outcomes

0.91
Avg Recall Quality
94%
Tool-Call Success
320ms
Voice Latency P50
11×
Memory Compression

Highlights

  • Private by design
  • Streams across web, desktop, mobile
  • Pluggable tools via JSON-schema
  • Works offline with local models
Discuss a Project Like This