Automating
the boring.

Scaling the interesting.

I build the piece of software a business is usually missing: the one that reads what customers send (lists, orders, forms) so a person doesn't have to retype it. Every build comes with a test that shows how often it's right, so you know what it gets wrong before you trust it.

1Receipts.Proof it runsa real system, live since july 2026

This is a system I built for cleaning, security and facilities companies in the UK. Every morning at 6:30 it wakes up on its own, reads every new government contract notice, works out which ones fit each company, and writes them a short email. No one presses a button. The numbers below come straight from its own logs.

P-01 · LINE-01 · S/N 0075
Running · last run 7 Sep, 10:39 UTC
78
78 mornings it has run by itself since 3 July 2026.
one bar per morning, last 40 · green ran, red failed
hover a bar for its date

It broke once, for about three weeks, when a data-format change silently killed the morning run. Fixed, and a failed run now leaves a visible trace. I leave the red stretch on the page because a system that never fails is a system nobody is watching.

P-01 · READContracts read
350
350 contract notices read and sorted so far. A person would have opened each one.
P-01 · CLIENTSCompanies served
7
7 companies each get their own matches, in their own sector and region.
P-01 · MODE
Human needed
BY HANDON ITS OWN
Zero manual steps from start to sent email.
2Origin story.What I do, and where I learned it
P-02 · OPERATOR

I automate the repetitive step between
“a customer sent us something” and “we acted on it.”

Lists, orders, postings, tenders: the part a person re-keys because nothing reads it for them.

I spent a year automating lead-finding for a sales-software company, got good at it, got bored of it, and noticed the same missing piece everywhere: nobody measures whether the automation is right. So now I build the thing, and the test for the thing.

What you hand meA process your team repeats by hand every week, and the tools you already pay for.
What you get backA system that does it on its own, documentation a new hire can read, and me on call to keep it running.
How I keep it honestI test every AI step against hand-checked examples before it touches real data, and I report the real score, not the flattering one.

ToolsMostly Python, plus whatever your business already runs on. Each project below names its tools.

P-02 · LOG
Where I’ve done this
Precise LeadsNov 2025 → now
GTM automation engineer (sales-tools automation)

A company that finds sales leads for other businesses. I automated how each lead gets researched, filled in with missing details, and added to the customer database.

8h → 2h
enrichment work per week, measured
Caprae Capital PartnersJun → Dec 2025 · US, remote
AI / ML engineer (intern)

An investment firm. I built the system that reads resumes and ranks candidates for their recruiting team.

10h → 3h
resume screening per week
B.Tech Computer Science · MSIT Delhi · 2026full resume ↗
3The evidence.Things I’ve builtmost proven first · every one links to something real
P-03 · BAY-01
01 · live · the system from step 1, every morning
Tender Radar

Finds government contracts for cleaning, security and facilities companies, every morning, on its own.

Government contract notices are spread across several websites that alert tools don't read. This pulls from all of them, uses AI to decide which fit each company, checks the deadline is real, and emails a short list. Built alone, running unattended since July.

pythonplaywrightgeminigithub actions
78
mornings run on its own
7
companies get their own list each morning
P-03 · BAY-02
02 · finished · deploying sep 2026
leadlens

Reads a job posting and tells you whether the “AI” role is real, dressed up, or a scam.

I hand-checked 72 real postings first, then measured the AI against them. When it scored a suspiciously perfect 100%, I found my own test was leaking the answers, fixed it, and published the lower, honest score.

pythongeminipydantic
82%
of scams caught
0
real jobs wrongly flagged
P-03 · BAY-03
03 · at caprae capital · 2025
Resume screening, automated

Reads incoming resumes, ranks the candidates, and books the interviews.

Built for a recruiting team spending most of a day a week reading resumes by hand. The system reads them, scores them against the role, writes the shortlist and schedules the calls.

pythonembeddingsslack
10h → 3h
screening per week
5 → 1.5
days to a shortlist
P-03 · BAY-04
04 · try it yourself
PresentAI

Paste your notes, get a slide deck in under a minute.

Structures the story, writes the slides and speaker notes, exports to PowerPoint. Live, free to try.

next.jspostgrespython-pptx
2h → 15m
to a first deck
<60s
for the first draft
P-03 · BAY-05
05 · at precise leads · in use
Lead enrichment on autopilot

Turns a raw list of names into CRM-ready leads with emails, company details and a score, automatically.

Pulls from several data sources, fills the gaps, checks the data, scores each lead, syncs to the CRM and posts a Slack alert when a good one lands. Enrichment used to be a person with a spreadsheet.

pythonclayn8ncrm apis
8h → 2h
enrichment work per week, measured
P-03 · BAY-06
06 · internal tool
Research-grounded writing

Writes a short, factual email about a company using only what a research step actually found.

AI left alone invents flattering details. Here, a research step runs first, the writer is only allowed to use what it found, and a separate checker reads every claim against the research.

claude codetypescriptweb research
95.5%
of claims checked true, 300-email sample
0
made-up facts found
P-03 · MONITOR
10-second loop · demo data
P-03 · CAPTION
Watch one
do the work.

A short video of an automated task board. The cursor you see is the system, not a person. It picks up each job, handles it, and flags the one that needs a human. Demo data.

4What broke.What broke, and what changed4 real entries · dated from the projects' own records

Every system on this page has failed at least once. This is the list. I keep it public because a person who tells you when it breaks is worth more than a system that claims it never does.

P-04 · INCIDENT.LOG
whensystemwhat brokewhat changed
2026-07tender radar

A data-format change silently stopped the morning run. Nobody noticed for about three weeks.

→ Fixed it. A failed run now leaves a visible trace instead of dying quietly.

2026-08demo pipeline

Rendering videos in parallel cut one down to 39 seconds over an 80-second narration.

→ One at a time now, with a length check before anything is sent.

2026-07tender radar

Short keywords matched inside longer words (“guard” inside “safeguarding”), so the wrong tenders got tagged.

→ Whole-word matching. Those exact cases live in the test set now.

2026-06leadlens

My own notes leaked the answers into the test, so the judge scored a perfect 100%.

→ Stripped them, re-ran, published the lower honest number.

5Field notes.How I think1 post · more as the work ships
6The call.Work with me
P-06 · DISPATCH
Twenty minutes.
No slides, no fluff.

Tell me about one process your team repeats by hand. I will tell you what I would automate first, what I would leave alone, and what it would take to keep it running. Free, and useful even if we never speak again.

P-06 · DOORS
If you're hiring
Need someone like this on the team?Open to full-time and contract, remote. Everything above is the interview prep.
If you run a business
Want the system built and looked after?First step is a fixed-price pilot on a month of your real requests, quoted before any work starts. During it, every miss is logged and you see the list, not just a score. Then a monthly retainer to keep it running.
room for 2 projects this quarter
replies within a day