Autonomy, with brakes.
Most automation projects fail at the last mile: permissions, exceptions, approvals, and the person who has to trust the thing on a Monday morning. Our approach is built around that last mile.
Principles
Six rules we
don’t break.
Start with the work, not the model.
We map the trigger, the handoffs, the exceptions, and the cost per run before we open a model console. If a rule works, we use a rule.
One workflow, one owner, one metric.
A pilot with three owners and no baseline is a science project. We name the metric before we build and measure it after.
Human where it matters.
Money, promises, legal, and people decisions get a checkpoint. The agent can draft, recommend, and prepare; a human commits.
Evals before go-live.
A golden set of real cases, including the ugly ones. We measure accuracy, cost per run, latency, and escalation rate. Then we watch them.
You own the system.
Code, prompts, credentials, runbooks, and documentation live in your accounts. No black box, no hostage pricing, no “call the agency” for a config change.
We’ll tell you when not to build.
Some workflows should be fixed with a checklist, a form, or a difficult conversation. We’d rather lose the project than ship a liability.
Reference architecture
A small system.
Not a platform.
Four layers, one workflow. We use the simplest stack that can run reliably in your environment — and we document every moving part.
Capture
The channels where work already arrives.
Decide
Rules for the common case; models for language and ambiguity.
Act
Read and write through scoped APIs, never shared passwords.
Guard
Approvals, evals, logs, and a fallback that ends with a human.
n8n or custom Node/Python · Postgres/Supabase · OpenAI / Anthropic / Google · WhatsApp Cloud API · Discord / Telegram / Signal bridge · your cloud or ours
Route by task: cheap models for extraction, stronger models for judgement. Every prompt versioned. No training on your data.
Every run logged with inputs, tool calls, outputs, latency, cost, and approval. Replay any failure.
WhatsApp/Meta message charges and AI model API usage are external and not included. Discord, Signal, and Telegram have no per-message platform fee, but hosting or bridge costs may apply. You pay these directly or at actuals with approval.
Security & data
DPDP-aware,
by default.
Agentic systems touch real customer data. We treat that as a design constraint, not a policy PDF.
We act as a data processor for client data and sign a DPA before kickoff. For our own website enquiries, we are the data fiduciary. This page is a summary, not legal advice.
Read the security & governance pagePurpose limitation
We access only the data the workflow needs, for the purpose you specified. No secondary use, no model training.
Least privilege
Scoped tokens, separate environments, revocable access. Production credentials never sit in prompts or spreadsheets.
Minimisation & retention
We redact what isn’t needed, set retention windows, and delete working data at handover unless you ask us to keep it.
Hosting choice
India-region hosting or your cloud account where residency matters. We’ll document every sub-processor.
Incident response
A named contact, a written escalation path, and notification timelines in the contract. We test the rollback before go-live.
Working together
What we need
from you.
The good-fit version
- A process owner
- One person who knows how the work actually gets done and can make decisions.
- API access
- Or a willingness to get it. We can work with exports and email for a pilot, but not forever.
- Two hours a week
- For review, edge cases, and approvals during the pilot. Less than that and we’re guessing.
- A baseline
- Current volume, hours, or error rate. If you don’t have it, we’ll measure it in the audit.
Not a good fit
“Just build a chatbot that does everything. We’ll figure out the process later.”
We’ve tried. It ends with a demo nobody uses and a team that trusts automation less than before.
Also not a fit: 100% autonomy on high-stakes decisions, no access to systems, or a race to the lowest hourly rate. We’re a small studio; we’d rather do fewer things properly.
Timeline
From audit to
running system.
| Stage | What happens | You get | Time |
|---|---|---|---|
| Audit | Shadow the process, interview the team, measure volume and exceptions. | Workflow map, cost model, ranked roadmap. | 1 week |
| Pilot | Build one workflow in production with real data and a small user group. | Working agent, approvals, logs, eval results. | 3 weeks |
| Harden | Break it on purpose; add fallbacks, cost caps, and documentation. | Runbook, training, rollback plan. | 3–5 days |
| Run | Monitor runs, fix edge cases, re-score ROI every quarter. | Monthly care plan and a named operator. | Ongoing |
Timelines assume API access and one process owner. We’ll give you a real date after the audit, not before.