Services
Advisory that comes with working systems.
Strategy where the decision is genuinely a decision. Engineering where something has to run. Both from the same firm, so the advice is answerable for the build.
What you can engage us for
Six lines. Most engagements combine two or three — a strategy question usually turns out to have a build inside it, and a build always turns out to have a policy question inside it.
Strategy Advisory
Executive roadmaps, operating models, vendor evaluations, buy-vs-build.
- 90-day AI action plan and KPI tree
- Model and vendor comparisons with decision memos
- Risk and policy baseline, DPDP/GDPR aware
Includes the recommendation not to build, when that's the answer.
Applied AI Engineering
Agents, copilots, RAG and content pipelines — built to run in production, then handed over.
- Task copilots, chat over your documents, content pipelines
- Agent systems on your own infrastructure and channels
- Eval harness — accuracy, edit-rate, cost
- Handoff docs, export packages and training
Proof: Smriti, BizClaw, CTO Dashboard
Knowledge & Content Ops
SOPs, promptbooks, editorial QA, and full content operations — planning, approval, production and multi-platform publishing.
- Knowledge audit and structure
- Editorial style and QA rules
- Verification gates on factual claims
- Content calendar, approval flow and scheduled publishing
- Bilingual packs as needed
Proof: Movisvami, AdCraftAI
GRC-Lite
Policy, risk register, access model — pragmatic guardrails that survive contact with a real deployment.
- Data handling and redaction rules
- Role-based access, least privilege
- Concise incident playbook
Enforced in architecture, not just documented. See below.
Enablement & Training
C-suite briefings, team workshops, playbooks, office hours.
- Executive seminars
- Hands-on team workshops
- Promptbook and sandbox
Delivered by someone who ships the systems, not a trainer reading slides.
Research Desk
Sourced landscape scans, synthesis memos and decision support.
- Competitive and vendor research
- Rapid synthesis with citations
- One-page decision briefs
Backed by 243+ published essays and a live release tracker.
How engagements work
Three shapes. The commercial terms are the same across all of them: IP in deliverables transfers to you on full payment per SOW; pre-existing IP and third-party components remain under their own licenses.
Advisory Retainer
Monthly access with standing hours. Fast feedback, reviews, and someone to think with before the decision is made.
- Response SLA
- Same day
- Turnaround
- 2–5 business days
- Cadence
- Quarterly roadmap refresh
Project SOW
Scoped deliverables with milestones and acceptance criteria agreed before work starts.
- Includes
- Design doc, evals, handoff
- Also covers
- Risk and data-handling notes
- Closes with
- Training session and package
Workshop
One day or a two-week sprint, for enablement or discovery — when the useful thing is momentum rather than a contract.
- You keep
- Templates and promptbooks
- Format
- Live exercises, hands-on
- Ends with
- Follow-up actions, owned
How you'll know it worked
Agreed before the work starts, measured after — because "the team feels more productive" is not a result anyone can act on.
Edit-rate is the honest metric.
How much of the AI's output a human had to fix before it could be used. It is the one number that cannot be gamed by a good demo, and it is the number most AI vendors will not quote you.
A copilot with a 60% edit-rate is costing your team time, whatever the accuracy score says. One at 10% is doing the job. We measure it at the start, and again at handover.
Every engagement ships with an eval harness so you can keep measuring after we've gone — which is the part that matters, since models change under you.
The eval harness
- Accuracy
- Against a task-specific test set your team agrees is representative — not a public benchmark.
- Edit-rate
- Proportion of output a human changed before use. Tracked over time, not sampled once.
- Cost
- Per task and per month, visible rather than buried. Multi-model routing means the cheapest adequate model does each job.
Skillify — the method behind the reliability
Named, because a method you can name is a method you can hold us to. Every capability we ship is checked against the same ten-step contract before it goes near your data.
The agent runs the script. It doesn't reason.
The failure mode of a business AI system is almost never a spectacular hallucination. It's the assistant that had a perfectly good tool available and decided to work the answer out instead — confidently, plausibly, and differently each time.
So the second item on the checklist is deterministic code: if code can do it reliably, the model is not permitted to. Dates come from the system clock. Figures come from a parser. Records come from a query. The model does what it's genuinely good at — judgment, synthesis and prose — and nothing else.
The rest of the checklist covers the written contract, unit tests, evaluations, and a maintenance ritual that runs on a schedule rather than when someone remembers.
What the checklist requires
- A written contract
- Name, trigger conditions, success criteria, and the hard rules — “always do X before Y”.
- Deterministic code
- Same input, same output. No reasoning where a script will do.
- Tests and evaluations
- Unit tests on the deterministic parts; evaluations on the judgment parts, covering the wrong behaviour as well as the right one.
- A compliance scorecard
- An installable check measuring whether the agent actually calls the scripts. Discipline that isn't measured decays.
Open foundations, and the exit
The two questions worth asking any AI firm, answered before you ask them.
- Built on open source.The platform underneath our work is MIT-licensed, and we contribute back — including changes merged upstream into projects we don't control. You can inspect what your system is made of.
- We release our own work too.Capabilities ship as installable packages rather than trapped in a private codebase, and several are public.
- No reselling, no referral fees.We don't take a margin on anyone's licences, which is why we have no reason to recommend the expensive option.
- Model-agnostic by design.Systems route across several providers by task. If one becomes worse or dearer, that's a configuration change, not a migration.
How you leave
Built before anyone asked for it — which is the only time it's credible.
- Your agent's memory and specialist configuration export with a single command
- The export imports into another installation, ours or anyone's
- Deployments run on infrastructure you own wherever the problem allows
- IP in deliverables transfers to you on full payment
- Documentation, packages and handoff prompts ship with the work, not after a request
Your data
The policy commitments, and the architecture that makes them true rather than aspirational.
- No client data trains public models. Ever. Where a hosted model is used, it's on terms that exclude training, and we'll show you which.
- Least-privilege access. Systems get the narrowest access that lets them work, and nothing adjacent.
- Redaction on request. Sensitive fields stripped before anything leaves your boundary.
- DPDP and GDPR aware. Not a compliance certification — a working literacy in what the regimes require, built into the design rather than audited afterwards.
- Self-hosting is the default option. Where it fits, the system runs on infrastructure you own. Your data never leaves.
Enforced, not promised
From systems running in production today. This is what "guardrails" means when it isn't a slide.
- Agents run in hardware-isolated VMs — only explicitly mounted directories are visible
- The application root mounts read-only, so an agent cannot rewrite the host and escape on restart
- The mount allowlist lives outside the project and is never mounted — an agent cannot edit its own permissions
- Credential paths blocked by default: SSH keys, cloud credentials,
.env, private keys - Symlinks resolved before validation, so traversal doesn't work
- Sessions isolated per group — one agent cannot read another's history
- Where an agent only needs to read, it gets a read-only database handle and is physically unable to write
Not ready to talk yet?
Start with the free 5-Fuse Business Triage.
A short self-diagnosis for founders and SME leaders: five business fuses, one priority, and a 90-day action plan. Open AI Insights and choose “Business in trouble?” to begin.
Try the 5-Fuse triageQuestions we get asked
Do you build full systems, or just advise?
Both, and the same people do both. We build production systems and run them — an agent platform serving six agents, a community management system live in a real building, dashboards a client extended himself. What we don't do is become a permanent dependency: every engagement ends with export scripts, documentation and a handover package, so your team or your partner owns it. If you only want a prototype to test an idea, say so and we'll scope one.
How do you handle our data?
Least-privilege access, redaction where you need it, and no training of public models on your data. Where it fits the problem, we deploy on infrastructure you own so the data never leaves your boundary at all. The security section above lists the mechanisms rather than the intentions — read-only mounts, allowlists held outside the agent's reach, blocked credential paths.
What does success look like?
Outcomes defined before we start, a measurable reduction in edit-rate, and decisions made on evidence rather than vendor claims. You keep the eval harness, so you can prove it to yourself after we've left — and catch it when a model update quietly degrades something that used to work.
Who owns the IP?
You do. IP in the deliverables transfers to you on full payment per the SOW. Pre-existing IP stays with its owner and third-party components remain under their own licenses — which we'll list, so there are no surprises in a due-diligence review later.
Which AI models do you use?
Whichever is right for the task, and we'll tell you why. Our own systems route across several models — one for orchestration, others for analysis and summarisation — so cost is optimised per task rather than per vendor relationship. We're not a reseller for anyone, and we have no incentive to recommend the expensive option.
Are we locked in to you?
No, and we'd rather prove it than promise it. The platform is built on MIT-licensed open source. Your agent's memory and configuration export with one command and import into another installation. Where the problem allows, the whole thing runs on hardware you own. IP transfers on payment. We built the exit before anyone asked for it, which is the only point at which building it means anything.
How big is the firm?
Small, deliberately. A founder-principal and a few others. Your first conversation is with the principal, and he stays involved in the judgment calls. If you need a hundred consultants, we're the wrong firm — and we'll say so on the call rather than three weeks into a proposal.
Book a call
Twenty minutes, no deck.
Tell us what's actually breaking. You'll get an honest read on whether it's worth building — including when the answer is that it isn't.
First conversations are with Lakshmi Narayana U, Founder & Principal Consultant. Hyderabad, India — remote-first, working globally.