Opus

Mobile companion

Operating Flow — Android

First-class Android client — not a companion. Every capability from web is available in your pocket: chat, calls, meetings with AI summaries, task triage, ask-anything. Optimized for one-handed use, glanceable, ambient.

41 screens7 categoriesEvery screen theme-awareLive, not mocked images

Stack

• Flutter · Impeller

• flutter_webrtc (LiveKit)

• firebase_messaging (FCM v1)

• Drift + PowerSync (offline)

• Passkey (WebAuthn)

Onboarding & auth7 Home & inbox4 Communication6 Meetings & calls6 Execution6 AI6 You & settings6

First 60 seconds

Onboarding & auth

7 screens

Operating Flow

Cognitive OS

by Ask Shiva Labs

Splash

Boot

9:41

Chat

Memory

Tasks

One place.
Everything you build.

Consolidate your team chat, meeting memory, and AI copilots into a single cognitive OS.

Welcome · 1 of 3

The pitch

9:41

Rahul

#askshiva-memory

HNSW is fine to 10M rows. We should stop rewriting pgvector and ship.

AI captured

Decision: Keep pgvector + HNSW, no rewrite

AI hears every
conversation.

Decisions, ideas, and tasks are captured automatically. Nothing gets lost in Slack scroll.

Welcome · 2 of 3

AI capturing

9:41

Your IP stays yours.

Ask Shiva Labs built Operating Flow so you never have to ship your model, data, or decisions to someone else's cloud.

Encrypted at rest

AES-256, per-org keys

In-house models only

SULTAN LLM runs on your infra

External providers OFF

No OpenAI, no Anthropic, by default

Welcome · 3 of 3

Security promise

9:41

Ask Shiva Labs

Work email

Password

New here?

Email + passkey

9:41

Create your organization

Free while in early access

Your name

Work email

Organization name

Password

Your private cognitive OS

AI captures every decision

External LLMs off by default

Already have one?

Create org

New tenant

9:41

Enable passkey for faster sign-in

Passkeys are phishing-resistant and never leave this device. Sign in with your fingerprint next time.

Bound to this device

No passwords to phish

Works offline

Enable passkey

Post-sign-in nudge

The daily hit

Home & inbox

4 screens

9:41

Good morning,Opus

Your empire is healthy

SULTAN v0.7 training on pace. Aventra ready to close.

5 projects8 people3 need review

Needs your attention

See all

HIGH@Opus in #askshiva-memory

Rahul mentioned you: "please review the memory refactor PR"

HIGHAventra replied

Anaya from Aventra sent NDA revisions — needs review before Monday.

AI captured today

Decisions

Ideas

Tasks

Live activity

Open chat

Opus#general

Capture

Home

Chat

Ask

Work

You

Home

Founder home

9:41

Notifications

HIGH@Opus in #askshiva-memory

Rahul mentioned you: "please review the memory refactor PR"

2h ago

LOWAskShiva Memory Review — summary ready

3 decisions, 5 tasks, 2 ideas extracted. 1 requires your review.

2h ago

MEDTask assigned: Set up Aventra repo + CI

Due today. Auto-generated by workflow.

2h ago

HIGHAventra replied

Anaya from Aventra sent NDA revisions — needs review before Monday.

2h ago

HIGHQ3 forecast due Friday

Neha needs your input on Q4 hiring plan.

2h ago

MEDKickoff — Aventra tomorrow 10:00

AI brief prepared. Review before meeting.

2h ago

Home

Chat

Ask

Work

You

Notifications

Prioritized

9:41

Alert

HIGH PRIORITYAskShiva AI

@Opus in #askshiva-memory

Rahul mentioned you: "please review the memory refactor PR"

Source

Opus

#general·12 min ago

Shiva suggests: Loop in Rahul and schedule a 15-min sync. Two similar issues resolved by lowering top_k.

Notification detail

One alert

9:41

Your day

Thu · Jul 10

Daily digest

Here's your Thursday

SULTAN eval is 2 days ahead of curve — v0.7 could ship Friday.
Aventra contract still pending signature — nudge them today.
Dev is overloaded — 6 open tasks across 3 projects.

Meetings today

AskShiva Memory Architecture Review

2:00 PM·AskShiva AI

???

SULTAN v0.7 Training Sync

2:00 PM·SULTAN LLM

???

Priority tasks

12 total

Add owner_type column to embeddings + backfill

AskShiva AI·HIGH

Gate rerank behind org_config.rerank_threshold

AskShiva AI·HIGH

Add rerank flag to eval baseline

AskShiva AI·MED

Daily digest

AI morning brief

Chat + DMs + threads

Communication

6 screens

9:41

Chat

Channels

DMs

Threads

general9:05 AM

Sam: closed Aventra — 6-month contract, $84k.

AI capturing

askshiva-devASKSHIVA3d

—

AI capturing

askshiva-memoryASKSHIVA9:57 AM

Rahul: dedup on (org, content_hash) at write time. cheap.

AI capturing

godengineGODENGINEYesterday

Arjun: TEE spike results in the doc

AI capturing

sultan-trainingSULTAN11:31 AM

Priya: MMLU 58.3 → 61.2. Ahead of the curve.

AI capturing

gesture-cricketGESTURETue

Kavya: new physics tuning spec attached

AI capturing

clientsCUSTOMAI32m

Sam: Aventra kickoff Monday 10am

AI capturing

leadership1d

Neha: Q3 burn came in $6k under

AI capturing

confidential3d

🔒 End-to-end encrypted — AI cannot read

Home

Chat

Ask

Work

You

Channels

All conversations

9:41

Chat

Channels

DMs

Threads

Rahul Verma2m

Lead Engineer (Backend)

yeah, I'll push the memory refactor tonight

Priya Nair18m

ML Engineer

training loss curve looking clean now

Sam Okonkwo1h

Growth & Sales

Aventra wants a quote by Fri

Kavya Iyer3h

Product Designer

sending the cockpit v3 mocks in a bit

Neha Bhatt1d

Ops & Finance

Q3 burn came in $6k under

Home

Chat

Ask

Work

You

Direct messages

1:1 with the team

9:41

askshiva-memoryASKSHIVA

4 members • active

AI is capturing — decisions, tasks & ideas

Rahul9:41 AM

Reviewing the memory retrieval fallback — the reranker is doing more harm than good below top-3.

Priya9:43 AM

Same. I ran evals last night — turning off rerank on <3 hits moves recall@10 from 61 → 74%.

🎯3

Opus9:45 AM

ok let's ship it. Two things: (1) gate it behind org_config.rerank_threshold, (2) add it to the eval baseline so we notice regressions.

Arjun9:48 AM

Related thought — we should probably move from pgvector to a dedicated store past the 5M row mark. Pinecone/Qdrant/Weaviate.

Rahul9:49 AM

@arjun defer that. HNSW is fine to ~10M rows. When we cross it we swap behind RagService interface — no rewrite.

AI captured: architectural decision to keep pgvector until 10M rows

Message #askshiva-memory

Channel view

AI capturing live

9:41

Rahul

online

quick one — did you get a chance to look at the memory refactor PR?

9:38 AM

skimmed it. Overall shape is right. One nit on the RagService boundary.

9:39 AM

yeah, I'll clean that up before merge. Rerank threshold flag going in this evening.

9:41 AM

nice. Also add it to the eval baseline — Priya asked. Otherwise LGTM.

9:42 AM

yeah, I'll push the memory refactor tonight

9:43 AM

Message Rahul

DM view

Rahul thread

9:41

Thread

12 replies • #askshiva-memory

Arjun9:48 AMOriginal

Related thought — we should probably move from pgvector to a dedicated store past the 5M row mark. Pinecone/Qdrant/Weaviate.

4 replies

Rahul9:49 AM

@arjun defer that. HNSW is fine to ~10M rows. When we cross it we swap behind RagService interface — no rewrite.

Priya9:50 AM

agreed with Rahul. Right now the reranker matters more than the store.

Opus9:52 AM

decision: keep pgvector. Add it to the decision log so nobody re-litigates.

Arjun9:53 AM

fair. Filing it away.

Reply in thread

Reply thread

Focused sub-thread

9:41

askshiva-memoryASKSHIVA

4 members • active

Rahul9:41 AM

Reviewing the memory retrieval fallback — the reranker is doing more harm than good below top-3.

Priya9:43 AM

Same. I ran evals last night — turning off rerank on <3 hits moves recall@10 from 61 → 74%.

Opus9:45 AM

ok let's ship it. Two things: (1) gate it behind org_config.rerank_threshold, (2) add it to the eval baseline so we notice regressions.

Arjun9:48 AM

Related thought — we should probably move from pgvector to a dedicated store past the 5M row mark. Pinecone/Qdrant/Weaviate.

Reply in thread

Copy text

Pin to channel

Delete

Message actions

Long-press sheet

Voice, video, screen share

Meetings & calls

6 screens

9:41

Meetings

6 recorded · 1 scheduled

AllTodayThis weekRecordedScheduled

AskShiva Memory Architecture Review

Today • 10:00 · 48m

ORVPNAM

Decided to gate rerank below top-3, delay pgvector→dedicated-store migration until 10M vectors, and add owner_type to embeddings now to prepare for per-user memory shadows.

3 decisions5 tasks2 ideas

SULTAN v0.7 Training Sync

Yesterday • 15:30 · 32m

PNAMO

Training is on-target for month-end. Set up int8 quantization benchmark to reduce inference cost by ~50%. Kicked off eval regression tracker.

2 decisions3 tasks1 ideas

Aventra — Discovery Call

Yesterday • 11:00 · 54m

OSO

Client wants a private LLM-based document ingestion for legal review. 6-month engagement. Kickoff Monday. Custom terms negotiated.

4 decisions8 tasks2 ideas

GodEngine Weekly

2 days ago • 09:00 · 41m

ORVAM

Constitutional gate now blocks 100% of external-provider calls on confidential/secret data. Audit chain verified. Next: TEE research spike.

1 decisions2 tasks3 ideas

Founders Deep Dive — Q4 strategy

3 days ago • 16:00 · 1h 22m

Focus split: 60% AskShiva/GodEngine (product), 25% SULTAN (moat), 15% client work (revenue). No new hires until $30k MRR sustained.

5 decisions4 tasks6 ideas

Start

Home

Chat

Ask

Work

You

Meetings

History

9:41

Meeting

AskShiva Memory Architecture Review

Today • 10:00 · 48m

ORVPNAM

4 participants

AI Summary

Decided to gate rerank below top-3, delay pgvector to dedicated-store migration until 10M vectors, and add owner_type to embeddings now to prepare for per-user memory shadows.

Rahul argued the RagService interface makes future migration cheap; Priya confirmed evals show rerank hurts below top-3. Opus locked owner_type tagging as non-negotiable for shadow readiness.

Keep pgvector until the corpus crosses 10M rows

OOpus·94% conf

Meeting detail

Recording + summary

Live capture: 1 decision proposed

REC

Opus

You

Rahul

Priya

Arjun

Live call

2×2 video

Incoming call

Rahul Verma

Incoming call · Ops Team

AI Brief

Recent thread: memory rerank fix. Rahul agreed to push refactor tonight; likely wants sign-off on org_config.rerank_threshold gate.

Decline

Incoming call

AI brief before ring

Extracting: architecture diagram detected

rag_service.py · askshiva/memory

1class RagService:

2 def retrieve(self, query, k=10):

3 vecs = self.embed(query)

4 hits = self.index.search(vecs, k)

5 if len(hits) < 3:

6 return hits # skip rerank

7 return self.reranker(query, hits)

8# gated by org_config.rerank_threshold

◐ auto-diagram

OPNAM

Rahul is sharing

14:22

Screen share

Live extract

9:41

Meeting ended

48 minutes captured

AskShiva Memory Review

Memory Review in progress

Summary

Generating executive summary…

62%

Decisions found

Keep pgvector, gate rerank, add owner_type

Tasks generated

Routed to Priya, Rahul, Arjun

Ideas captured

Debate-room reasoning nodes, brief drafts

Meeting ended

48 min captured

Ideas, tasks, goals

Execution

6 screens

9:41

Work

8 tasks assigned across 4 projects

To do (5)In progress (2)Blocked (1)Done

ASKSHIVA AI

HIGH

Gate rerank behind org_config.rerank_threshold

RVRahul

Today

ASKSHIVA AI

MEDIUM

Add rerank flag to eval baseline

PNPriya

Thu

CUSTOMAI

URGENT

Set up Aventra repo + CI

RVRahul

Today

GODENGINE

LOW

TEE research spike (Intel TDX vs AWS Nitro)

AMArjun

Next week

GODENGINE AI

LOW

Add TEE mode toggle to GodEngine adapter

AMArjun

2 weeks

Add task

Home

Chat

Ask

Work

You

Task board

Kanban on mobile

9:41

Task

ASKSHIVA Proposed by AI88%

Gate rerank behind org_config.rerank_threshold

Introduce an org_config.rerank_threshold knob so the reranker only runs above N hits. Priya's eval shows recall@10 moves 61 → 74% when we skip rerank on <3 hits. Ship gated, add to eval baseline.

Assignee

Rahul Verma

Lead Engineer (Backend)

Due

Today

Priority

HIGH

Source conversation

#askshiva-memory · 9:45 AM

"let's ship it. Gate it behind org_config…"

Activity

Priya Nair12m

Adding eval baseline PR now — should be ready before you start.

Rahul Verma4m

Cool. I'll wire the flag through RagService and land the PR today.

Task detail

One task, deep

9:41

AI proposed6

6 tasks extracted from today's meetings need your review

Auto-assigned by AI. Accept to route, edit to tweak, or reject to discard.

AI91%

Add owner_type column to embeddings + backfill

ASKSHIVA→PNPriyaAskShiva

AI90%

Benchmark int8 quantization for SULTAN v0.7

SULTAN→PNPriyaSULTAN

AI78%

Cache constitutional gate decisions by (dataClass, provider)

GODENGINE→AMArjunGodEngine

AI66%

Draft Aventra NDA revision response

CUSTOMAI→SOSamCustom

Home

Chat

Ask

Work

You

AI triage

Accept / edit / reject

9:41

Capture idea

AI shadow briefs could pre-warm the debate room too — pull the last 48h of contested threads into an agenda before we even open it.

Suggest projectSuggest clusterNovelty check

Similar to "AI shadow briefs" (72% novelty) — link as related?

Capture idea

Voice + text

9:41

Project Genome brainstorm

5 live · 12m

Opus9:12

What if every project had a living "genome" — mission, arch, contributors, decisions, risks — that updates itself from graph state?

Novelty

83%

Kavya Iyer9:14

Love it. A tree layout on desktop, tabbed on mobile. Genome should include a "what would break this project" section.

Novelty

61%

Rahul Verma9:16

Materialize it from decision events + task events. Read-through cache, invalidate on any node write in the project subgraph.

Novelty

70%

Synthesis AI· synthesis9:17

Synthesizing: 3 threads converge on a "self-writing project brief". Nearest neighbor: Living Docs Updater (v5 prompt, 340 calls). Suggest promoting to spec.

Novelty

88%

Priya Nair9:19

Add an "unresolved contradictions" section — every project probably has 2-3 lurking ones we forget about.

Novelty

91%

Brainstorm

Live room

9:41

Goals

4 active · Q4 2026

Company goal

Launch AskShiva V3 to internal users

Q4 202662%

Ship agent memory subsystem78%

Ship debate room MVP40%

Product goal

SULTAN LLM v0.7 shipped + serving

Q4 2026

PN74%

3 objectives · training on target

Revenue goal

Sustain $30k MRR from client work

Q4 2026 → Q1 2027

SO48%

New goal

Home

Chat

Ask

Work

You

Goals

Goal → Objective → Milestone

Ask, search, agents

AI

6 screens

9:41

Ask anything

Your entire company's memory

Try asking

When did we first discuss agent memory?Show every meeting where SULTAN was discussedFind every architecture decision made by RahulShow why we rejected RedisWhat are all unfinished ideas from last month?

Recent

Why did we choose pgvector over Pinecone?

12m ago

Summarize the Aventra kickoff prep

Yesterday

Ask about anything…

Home

Chat

Ask

Work

You

Ask anything

Company GPT entry

9:41

AskShiva

org-only · memory active

org

Why did we choose pgvector over Pinecone for memory?

You debated it Apr 27. Rahul argued pgvector kept ops surface small while under 10M rows. Priya wanted Pinecone for latency but the Redis migration burn was fresh. Decision: keep pgvector until 10M rows, revisit with a rerank benchmark.

Meeting · TodayDecision · Apr 27#askshiva-memory

Who owns the rerank benchmark?

Priya, with Arjun reviewing. Ticket TASK-812 is still open — no PR yet.

Task · TASK-812Priya Nair

Reply to AskShiva…

AI conversation

With citations

9:41

Show why we rejected the vector-store migration

AI answer

You rejected the vector-store migration in today's architecture review. Rahul argued pgvector's operational simplicity beats marginal latency wins under 10M rows. Revisit when a rerank benchmark exceeds the current bge baseline.

Meeting · TodayDecision · Rahul#askshiva-memory

DECISION

Rejected: migrate pgvector → dedicated vector store

Meeting · Memory Architecture Review · Today

Rahul: staying on pgvector under 10M rows; the ops surface is worth more than the marginal latency win. Pinecone deferred.

0.94

match

MESSAGE

Priya on rerank benchmark

#askshiva-memory · Priya Nair · 2h ago

"MMLU-style eval on rerank finished — bge-reranker still edges Cohere on our corpus. Sticking with local."

0.88

match

MEETING

Memory architecture — full transcript

Meeting · 42 min · 6 participants

Discussion covered vector store choice, rerank strategy, and eventual sharding threshold at 10M rows.

0.81

match

Home

Chat

Ask

Work

You

Semantic search

Meaning-first

9:41

AI brief · Aventra kickoff

Prepared by Meeting AI

Auto-brief

Meeting in 15 min

Tomorrow · 10:00 · 45 min · Aventra kickoff

RSO+ 2 from Aventra

What to know

•Contract signed yesterday. $84k, 12-week engagement.
•Their CTO Anaya cares about audit trail — flag GodEngine gate.
•NDA revisions still open — do not commit to timelines yet.

Recent context

Anaya · Aventra

"Sending revised NDA tonight — mainly IP carve-out language."

Sam

"Kickoff agenda approved. Milestones will slip if NDA isn't done Monday."

Decisions to make

1.Which model — SULTAN v0.7 vs AskShiva runtime for their pilot?
2.Onboard Rahul or Arjun as their engineering point of contact?

Suggested opener

"Anaya, before we jump in — the IP carve-out language in the NDA looked reasonable. Want to close that first so nothing blocks week one?"

Meeting AI brief

Pre-meeting prep

9:41

AI agents

11 agents · GodEngine gated

CEO AI

Founder's counsel

421

calls

94%

csat

Architecture AI

System designer

812

calls

91%

csat

Coding AI

Pair programmer

2,140

calls

88%

csat

PM AI

Autonomous project manager

645

calls

90%

csat

Research AI

Domain expert

402

calls

87%

csat

Memory AI

Second brain

5,820

calls

95%

csat

Home

Chat

Ask

Work

You

AI agents

Your specialists

9:41

Review AI captures

From: Memory Architecture Review

5 of 8 remaining

3 done

DECISION

Extraction AI · v12

Keep pgvector as the memory store until we cross 10M rows.

Rationale

Operational simplicity outweighs the marginal latency win under current load. Rahul referenced the Redis-migration burn from Q1.

Rejected alternatives

PineconeWeaviateRedis vectors

Confidence91%

Attributed to Rahul · 12 min ago

Swipe right to accept · left to reject

Insight review

HITL swipe deck

Profile, security, appearance

You & settings

6 screens

9:41

You

Opus

Founder & CEO

Ask Shiva Labs

214

decisions

812

tasks

meetings

AI shadow

92% coverage

92%

Trained on 4,812 interactions across the past 90 days

Home

Chat

Ask

Work

You

Home

Chat

Ask

Work

You

Profile

Your AI shadow

9:41

Settings

Personal

Organization

Security

Home

Chat

Ask

Work

You

Home

Chat

Ask

Work

You

Settings

Everything else

9:41

Security

Passkey enabled

Face ID + WebAuthn · registered 4 mo ago

Devices

MacBook Pro 16"

active now

this device

Pixel 8 Pro

2 hours ago

Ubuntu workstation

4 days ago

Audit chain

verified

18,412 entries · hash sealed

Home

Chat

Ask

Work

You

Home

Chat

Ask

Work

You

Security

Devices + audit

You're entering a
confidential room

AI cannot see this conversation. Messages are end-to-end encrypted between members only.

No summarization

Not added to knowledge graph

Auto-purges after 30 days

This action is logged in the audit chain.

Confidential room

AI-off warning

9:41

Appearance

Theme

Text size

Medium · default

Accent glow

Subtle halo around focused elements

Home

Chat

Ask

Work

You

Home

Chat

Ask

Work

You

Appearance

Theme picker

9:41

Your organizations

Ask Shiva Labs

current

Founder & CEO · 7 members

RPAK

Switch to

Signed in as opus@askshiva.ai

Home

Chat

Ask

Work

You

Home

Chat

Ask

Work

You

Org switcher

Multi-org

Ready to code these? Every screen is a plain React component right now.

Same layout translates to Flutter widgets 1:1 — mock data becomes real API calls, phone frame becomes actual device chrome.

Configure push