Work
Systems that declare their own limits.
A platform, the systems built on it, the agents running our own firm, and what we've put into the open. Four kinds of proof — separated, so we don't present them as the same kind.
The platform
What everything else runs on
6 agents · 5 channelsMost business AI is a subscription to somebody else's cloud. This runs on hardware you own, reachable from the chat apps your team already uses.
Agent platform · Self-hosted
BizClaw
A business AI assistant for WhatsApp and Telegram, built on NanoClaw (MIT). Each agent runs in its own hardware-isolated Linux VM on Apple Container — about a second to boot, no Docker daemon. Multi-model by task through OpenRouter: Claude orchestrates, Kimi analyses, Gemini summarises. A LoRA checkpoint fine-tuned on 👍/👎 feedback takes production traffic, with fallback. Capabilities install, update and roll back through a real skills engine — 60+ of them, from channel adapters to a one-command sales CRM.
Run the script, don't reason
Every skill ships against a ten-step checklist requiring deterministic code for anything code can do — and a separate installable scorecard measures whether the agent actually calls those scripts or quietly improvises. The security model works the same way: the project root mounts read-only so an agent can't rewrite the host and escape on restart, and the mount allowlist lives outside the project entirely, where no agent can edit its own permissions.
Running on it today
- Andy — personal assistant, WhatsApp
- Andy · chief of staff — Telegram; orchestrates Iris (content), Leo (revenue), Scout (intelligence); morning brief and delegation log
- Movisvami — cinema content pipeline, five platforms
- Smriti — apartment community agent, live in a real building
- NGMF Sales — daily collections report, 8pm IST, emailed to stakeholders
- Global brain — shared memory: status, observations, proposals, growth and delegation logs
Tier one
Engagements & products
7 buildsBuilt on real data, hardened after they worked, packaged so they could be deployed again without us in the room. Each has a full case study.
SME & association operations · on BizClaw
Smriti
A residential community's committee asks questions of its own records in WhatsApp — and gets answers, from an agent that cannot change those records.
Physically unable to write
The query path opens the database read-only. Not instructed not to edit — unable to. Verified against the live database, negative tests included.
Enterprise CXO visibility
CTO Dashboard
A demo the client extended himself — onto his own SharePoint, across a 119-project portfolio — and now runs without us.
Which statuses are inferred
Blocked and paused are explicit. At-risk is inferred from free text, and the dashboard says so before a CXO takes it into a board review.
Consumer product & ranking systems
WatchWise
A time-boxed recommendation engine: tell it how long you have and what you're in the mood for, get three ranked picks with the reasoning attached.
The weights are yours
All nine scoring components are published and adjustable — and every pick shows its opportunity cost against the best available one.
Creative & agency workflow
AdCraftAI
A creative operating system that refuses to write copy before it has a strategy — brief, strategy, concept, execution, testing, in that order.
Mockup, not final art
Generated visuals are labelled art direction, not production assets. A soft wall pauses on rate limits without losing a minute of work.
Knowledge platform & AI practice companion
AI Insights
A 243+ article archive turned into a working practice companion: ask questions across the writing, explore topics, save playbooks and bookmarks, or switch into a structured 5-Fuse business triage.
Answers show their sources — and their age
Semantic retrieval returns source citations with publication dates; voice commands drive the same React handlers as the visible controls rather than a separate hidden workflow.
Content operations & publishing
Movisvami
Plan a week, approve it day by day, produce the assets, publish across five platforms — all from a chat thread. Running our own cinema brand as its reference deployment.
A script decides what's real
Every quote is verified against real sources by code before it can be used. Fails the check and it's dropped — however plausible it looks.
SME finance & receivables
The Collections Agent
A daily receivables report assembled from the sales team's own chat group and emailed each evening — a number the business used to get monthly, if at all.
The numbers come from a parser
Code extracts the figures; the model only writes the commentary around them. It never does the arithmetic.
Tier two
Our own operation, automated
4 systemsThe honest test of an AI advisory is whether it runs on the thing it sells. These are the agents doing our work — a chief of staff, a weekly publication, the first hours of every engagement. Movisvami sits in tier one now because it became a product, but it started here, running our own brand.
Pipeline · Publishing
Director's Cut
The weekly newsletter, generated as email-client-safe HTML from a structured content dump — Medium pieces, Readwise highlights, GitHub activity, Movisvami picks — in full brand, with tracked links and Listmonk template functions.
Weekly issue. Table-based layout, all styles inlined, section rules encoded as a reusable skill.
Internal tool · Advisory delivery
Discovery Desk
Rough client discovery notes in, structured advisory report out — summary, phased use cases, a scored pilot readiness breakdown, risks and stakeholders, proposal, and a ready-to-send follow-up email. The first hours of every engagement, made repeatable.
Used on live engagements. Refuses to generate on thin input rather than inventing a report — and preserves the previous one when it does.
Agent · Chief of staff
Andy
Runs the firm's day from Telegram — orchestrating three specialists (content, revenue, intelligence) rather than doing their work. Delivers a morning brief off a shared memory of status, observations, proposals and world context.
It reports how it operated. The brief carries a delegation log — what it handed off and why — plus a Monday review of the week's decisions, self-corrections included.
Live artifact · Research
Claude Feature Tracker
A living page tracking model and feature releases, refreshed on a schedule rather than rebuilt by hand. The research desk behind a lot of the writing — and the reason the essays cite dates rather than impressions.
Continuously refreshed. Built as a live artifact, not a static document.
Tier three
Open work
6 projectsOur commercial platform is built on MIT-licensed open source, and we contribute back. Some of this was commissioned by nobody and built on weekends — labelled as such, because the distinction matters. But the foundation under the paid work is the same foundation, and that is the point of showing it.
Fork · AI video editing
OpenCut — approval-gated edit agent
An agent that inspects footage, transcribes it and proposes a complete edit plan. It cannot press render. Hash-verified provenance on every asset it touches.
Same rule that should hold for any agent near a send, publish or deploy button. Published write-up.
Contribution · Upstream merged
WorldMonitor — India & Cinema variants
Two new configurations and a custom deck.gl map layer covering 23 hubs, contributed back to the upstream project rather than kept as a private fork.
PR #5724 upstream · published article with original diagrams.
Field notes · Agent orchestration
Hyperagent fleet over MCP
Alpha testing a 40-agent fleet driven from another application entirely — relays, fan-outs, and a public Claude skill shipped from the work. The platform's quality loop caught its own agent fabricating; some flagged "fabrications" turned out to be true.
Two published essays. A system being wrong out loud is worth more than one that never admits doubt.
Substrate · Agent runtime
NanoClaw
The agent runtime Smriti is built on. Documentation reviewed and written up publicly — the open-source layer directly under a commercial deployment.
Published essay. The thing we contributed to is the thing we deployed.
MCP server · Research tooling
Competitive Intelligence MCP
A server exposing event search, brief generation, dashboard creation and saved-article analysis to any MCP client — built to make competitor research repeatable instead of ad hoc.
Running in production on our own research workflow.
Plugins & skills · Released
Cowork plugins & skills
Chief of Staff (briefings, meetings, planning, comms), Team Status Digest, a recent-research engine, and a knowledge wiki — packaged as installable plugins and skills rather than kept as personal prompts.
Installable packages. The same packaging discipline applied to client deliverables.
Two of these were built with no client, no brief and no deadline — and both independently arrived at the same rule as the paid work. OpenCut's agent can plan an entire edit but cannot render it. Hyperagent's quality loop flagged its own output as fabricated, and was sometimes wrong to.
That's the point of showing them. A principle you apply when a client is paying is a policy. A principle you apply when nobody's watching is a practice.
Start a conversation
Tell us what's actually breaking.
Not what you think the AI solution is — that part is usually wrong, including when we guess it. Describe the problem and you'll get an honest read on whether it's worth building.
First conversations are with Lakshmi Narayana U, Founder & Principal Consultant.