I automate business processes and prove the impact.

Every broker deal email meant about twenty minutes of reading, retyping and filing by hand. The record was only as good as the typist.

I built a Power Automate flow that uses AI extraction to pull the deal fields out of free text email. It creates the record, files the attachments, and blocks duplicates on a unique key.

Employer system, so no public link. I can walk through the live flow and its reconciliation report in an interview.

Maintenance and support requests came in as free text emails. No reference number, no queue, and no way to see what was still outstanding.

I built a flow where AI classifies the mail into structured fields, writes a numbered ticket with an audit trail, and auto replies to the requester with the number and a correction path.

Employer system. In an interview I can send a test mail and show it become a numbered ticket live.

Handover detail sat inside meeting recordings. Turning a transcript into a written, searchable record was a manual write up every single time.

I built it to the operations stakeholder's spec. AI drafts a structured handover record from the transcript, a human approves or edits it, and that approval automatically creates the maintenance entry.

Employer system. I can walk the full loop live, transcript to approved record to maintenance entry, in an interview.

I inherited five scheduled HR probation review flows. No monitoring, no runbook, and nobody but the original author could maintain them.

I took ownership, rebuilt the connections, and tested all five end to end against real review dates. Then I added failure alerting and wrote a runbook covering triggers, schedule and recovery.

Employer system. The runbook exists, and I can walk through all five schedules and their failure alerting.

Scanned site paperwork, mostly daily cashups and supplier invoices, was read, sorted and filed by hand at the branch.

I took ownership of the scanning pipeline and rebuilt it. AI Builder OCR reads each scan, classifies it into one of eight categories, and files it into the right shared library with a logged trail.

Employer system. I can walk through the live pipeline, its filing log, and the written rollout plan for the second site.

Bulk buyers ordered by phone, WhatsApp and walk in. Missed messages meant lost orders, and no order was recorded anywhere at all.

I built it end to end in three days. One link or QR code to order, an instant reference number, a mobile approval step for the branch manager, then an automatic confirmation and a prep sheet.

Employer system, still in business walkthrough, so I claim no value yet. I can demo a test order end to end.

New hire onboarding ran on two macro-enabled Excel workbooks. Desktop only, one person in the file at a time, and no view across employees.

I audited the live process and wrote down four failure modes before building anything. Then I replaced both workbooks with a web system: a 45 task workflow per hire, a cross employee dashboard, automated task and probation emails, and CSV export.

Employer system, in business demo, so I claim no hours saved yet. I can walk the workflow, the dashboard and a test mode email run in an interview.

Applying by hand meant reading every spec and rewriting my CV and cover letter each time. Hours per application, and the scoring had no tests.

I built it in Next.js with structured Claude outputs, so every score is comparable. It ships in a multi stage Docker image, with 212 unit tests and a GitHub Actions pipeline gating typecheck, lint, tests and build.

• See the real scoring output, all 21 listings
• The CI gates and their real output

Proof pages show real scoring output for all 21 listings, plus the gates run locally with their real exit codes.

Agentic AI is easy to claim on a CV and hard to prove. Job specs ask for LangGraph and I had nothing runnable to point at.

I built a LangGraph state graph: four specialist agents (classifier, extractor, validator, router) under a supervisor that owns retry, escalation and degradation. Two agents call the model, two are pure Python rules.

• See all 26 documents, four runs and the raw audit trail

The proof page comes straight from the run files. All 26 documents are synthetic, so accuracy is on data of that shape.

The source system keeps a status column and no history. Nobody can answer what fraction of deals reach offer, because yesterday's status was overwritten today.

I built a star schema warehouse on PostgreSQL: two Type 2 slowly changing dimensions so history reports as it was, 84 monthly partitions on the event fact, 51 data quality assertions, and six measured performance case studies.

• See the 10 queries and their real output
• Source on GitHub

Public repo. One command rebuilds the whole warehouse, and every query output, plan and buffer count is committed, so the numbers can be checked.

One person working sequentially by hand. Build, then research, then monitor, then apply, one long thread at a time.

I built a rig of custom subagents, MCP servers, hooks and scheduled agents. Agents run research and verification in parallel, hooks enforce my standing rules, and independent agents check the work before it reaches me.

Local setup, nothing to link. The proof pages on this site were generated by it, and the config and run logs are on disk.

Dataverse is the platform depth most Power Platform people skip. They stop at lists and never learn tables, relationships, security roles or model driven apps.

I specced the build and seeded synthetic data: Broker, Property, Deal and Intake Log tables, an alternate key for dedupe, business rules, two security roles with column security, and a model driven app.

Nothing is built yet, so nothing to open and no value claimed. The spec, three runbooks and seed CSVs exist and can be shared.

During the service rush, calls rang out unanswered. Bookings and takeaway orders were lost to a busy signal.

I built it on Retell and Twilio with an SA English and Afrikaans voice. It takes bookings and orders, n8n writes each one into Supabase over webhook, and the branch dashboard shows the day.

• receva.co.za

The pilot line belongs to a client, so the number stays private. Call recordings and transcripts can be walked through on request.

Cold outreach had nothing interactive to show. Just a claim in an email and nothing for the prospect to click.

I generate a per prospect page from their public business detail, and the Claude API answers real caller questions live in that business's voice. Six are deployed from a library of sixty five.

• Try the live AI receptionist desk
• Try the law desk
• Try the dental desk
• Try the recruitment desk
• receva.co.za

Nothing is scripted. Ask the desk anything and the answer is generated live.