{ ai automation studio }

Automation, second nature.

I’m Aasim. I build the AI systems that take repetitive work off your desk. n8n workflows, custom agents, data pipelines, wired into whatever your team already has open. They ship deployed and monitored, then they keep running without you.

See what I've built Start a project
● currently accepting clients scrolled: 000%

{ selected work }

Built, deployed,
still running.

Six systems, each with a case study and a working build behind it. Open the evidence, watch a run, or try the live demos.

01 ● live demo

Competitor
Intelligence

Name a competitor’s domain and eight agents read the whole company: positioning, pricing, SEO, content. Every gap they find gets ranked into a plan you can work through. The pipeline will not render a section without two citations from pages the crawler actually fetched, so thin signal comes back marked insufficient-data and stops there.

8 agents · 9 report sections · 2+ evidence per gap

  • multi-agent
  • web crawl
  • modal
View case study
02 ● published audit

SEO Decay
Audit

The old SEO decay showed up in clicks. The new kind shows up in citations: your page still sits at position three, and the AI answer above it already gave the reader what they came for. This scores 16 pages from four major content sites on both axes and sorts each into one of four verdicts, with every number labelled by where it came from.

16 pages · 4 sites · 64 prompts · $0/mo hosting

  • ai visibility
  • citation scoring
  • astro
View case study
03 ● live app

Search
Stack

Nine SEO modules in one platform you run yourself: rank tracking, site audit, keyword research, Search Console insights, a prompt explorer. One codebase builds both the hosted and the self-hosted version, and the self-host build drops the auth layer completely, so deploying it opens straight into the dashboard on your own Cloudflare account and your own DataForSEO key.

9 modules · multi-project · edge-deployed · bring your own keys

  • self-hosted
  • seo platform
  • cloudflare
View case study
04 ● live app

Customer
Intelligence

Support conversations come in as tickets and get clustered into named issues, then ranked by how much each one grew against the week before. Open any signal and the conversations that produced it are sitting right there. Knowledge gaps, suggested actions and the executive report all read off the same dataset, and nothing fires until a person approves it.

9 modules · issue clustering · evidence per signal · synthetic dataset

  • support ops
  • issue clustering
  • vercel
View case study
05 ● live app · demo login

AI Financial
OS

Cash flow, runway, invoices, recurring bills and alerts over one ledger. The forecast beside them is plain arithmetic: pick Base, Conservative or Growth and it prints every assumption it used, every projected line in date order, and the things it deliberately left out. The AI CFO talks about figures the system already calculated, names the period behind each one, and has no write access to anything.

9 modules · 3 forecast scenarios · deterministic projections · assumptions printed

  • cash forecasting
  • scenario modelling
  • vercel
View case study
06 ● live demo

Bid
Intelligence

BidOS takes an RFP in and gives back a proposal someone can sign off on. It parses the documents, anchors every requirement to the span it came from, scores the bid / no-bid call, then drafts only from approved evidence. A verifier reads each factual sentence against the chunk it cites, flags the ones nothing supports, turns the gaps into questions routed to an expert, and holds the export shut until they clear.

83 requirements · 13 evidence traps · 4 human gates · 16+ export checks

  • go
  • hybrid retrieval
  • claim verifier
View case study

{ how it becomes second nature }

Three steps.
Zero babysitting.

Traced below through the competitor intelligence build, the one at the top of this page.

  1. 01

    Map

    We walk through the process you’d rather never do by hand again: what sets it off, what it touches on the way through, and where it usually breaks. One call, thirty minutes.

    your part: thirty minutes on a call

    One job, scoped tight: read a competitor end to end and rank every gap. The failure mode was obvious from day one, which is a model inventing a price it never found anywhere.

    scope
  2. 02

    Build

    I build the whole thing: agents, integrations, retries, error handling. What you review at the end is a working build you can run yourself.

    your part: review a working build

    Eight agents, one per surface, all writing into a shared nine-section report. The pipeline holds the evidence rule so the model never gets a vote on it: two or more citations from pages actually crawled, or the section comes back as insufficient data.

    engineer
  3. 03

    Run

    It ships to Modal, Vercel or Render, fires on a schedule or a webhook, retries its own failures and reports back. You check in when you feel like checking in.

    your part: none

    Live on Modal right now. Name any competitor’s domain and the full pass runs on demand.

    autopilot

{ systems I build with }

The stack,
end to end.

01

AI

Model routing and structured extraction: a model chosen per task, so the cheap jobs run on cheap models and the bill stays legible.

OpenAI Claude Gemini OpenRouter

02

Automation

The orchestration layer: what fires, in what order, and what happens when a step fails at 3am.

n8n Trigger.dev Python AI Agents Webhooks

03

Data

Getting the inputs. Crawling sites, pulling APIs, and cleaning what comes back into something a model can actually work with.

Firecrawl Scraping DataForSEO Search Console APIs Data Pipelines

04

Workflow

Where the output lands: a task in your tracker or a row in your database. Dashboards get opened twice and then forgotten.

ClickUp Supabase PostgreSQL

05

Interface

When a system needs a front door. Typed routes, server rendering, and a UI that ships with the build, with no separate design phase in front of it.

React TanStack Start Tailwind daisyUI

06

Infrastructure

Deployment and uptime. Scheduled jobs, serverless functions, edge workers, long-running containers, and version control on every workflow I ship.

GitHub Modal Vercel Cloudflare Workers Render

{ the human }

Aasim,
behind the machines.

If a human does it twice, a machine should do it a third time.
freelance automation engineer

I’m Aasim, and Second Nature is just me. I built everything on this page.

The six builds above share one habit, and it’s the one I’d want you to judge me on. When they don’t know something, they say so. The competitor crawler only reports a price it can cite to a page it fetched, and everything else comes back marked insufficient-data. The financial forecast is arithmetic with no model anywhere near it, and it prints every assumption it used, down to the invoices it left out of the maths.

The bid demo is the clearest case. I wrote the sample RFP it answers, and I put 13 requirements in it that nothing in the evidence library can support, plus one hidden prompt-injection line. Part of that demo’s job is to fail in front of you: thirteen questions it has to decline, and an instruction it is supposed to ignore. A demo where everything works tells you nothing about the Tuesday when the data is bad.

Retries, error handling and monitoring go in with the first build, on the first invoice. Everything above is still up, and none of it needs me there.