Applied AI · Agents in Production · Enterprise Deployment

From model
to Monday morning.

I take AI models from chat to real solutions — systems that do actual work inside real businesses, with real money moving on their output. The most recent one applied ~$100M of cash in its first two days, designed, built, and shipped with the people who use it. Behind that: thirteen years of making new technology actually land in complex organizations. The model is rarely the bottleneck. Making it real is.

See the work Visit the lab
100M
Scroll to explore

01 — Work

Agents in production.

Build stories: a problem in, a working system out.

2026 · In production

The cash-application agent

A Fortune 500 media & streaming company had a classic AR problem at streaming scale: customers pay large lump sums decoupled from invoicing, and someone has to figure out what the money is for. I designed and built an LLM agent that reconciles those payments against open receivables across three lines of business — ~$100M of cash applied in its first two days.

I was the only technical person on the project: scoping with senior finance leaders, agent and conversation design, tooling, data pipeline, implementation, iteration in production.

  • Divide the labor
    Give the model only the work that needs judgment; everything computable is done by plain code it can call. An agent allowed to do arithmetic will eventually do it wrong — this one is forbidden to.
  • Design for doubt
    A system keeps trust by saying how sure it is, claim by claim — and by never being allowed to sound more certain than it is. Trust dies the first time it is confidently wrong.
  • Meet the data where it is
    Enterprise data never has the clean identifiers you wish for. Instead of letting the model guess across the gaps, it reads the results of exact, bounded search — it interprets evidence, it does not invent it.
  • Adoption by design
    People adopt automation that amplifies their authority and resist what replaces it. Built with the analysts, not around them: the agent drafts every email and sends none — they keep the decision and the byline.
~$100M
Cash applied · first two days
Client confidential by agreement
Summa · 2025
$5M

The AI-accelerated enterprise close

A six-month-stalled technical problem, reframed as two sequenced programs — first carving a Fortune 500 US industrials company out of its former parent's multi-country SAP estate, then an S/4HANA conversion. A custom AI diligence stack compressed validation from weeks to 48 hours and closed the $5M engagement: Summa's first Fortune 500 US reference.

Accenture → Summa
17 countries

Built it — then came back to sell it

Designed and built SUMMA, a 17-country shared-services center for a $7B multi-industry holding, consolidating Finance, Procurement, HR, and IT. Six years later I was brought back to commercialize what I'd built into a market-facing BPO — few people get to operate both ends of that arc.

Founder · 2020–2024
0 → 4,000

Multi-agent, before it was mainstream

Co-founded Gigflow, an EU-Innovation-Fund-backed AI startup (€300K grant) looping automation with LLM reasoning from the moment the OpenAI API opened. Zero to 4,000 monthly signups in five months, $4M valuation in year one, a 12-person team.

gigflow.co ↗
The foundation

Nine years at Accenture across two stints — ERP programs, post-merger integration, and operating-model design for global leaders in brewing, consumer goods, aviation, and mining. Then four years building Gigflow from zero — the founder chapter that turned an enterprise operator into an AI builder. B.S. Industrial Engineering, Universidad de los Andes. Working languages: English, Spanish, Portuguese, Polish.

02 — The Lab

Real businesses, run as experiments.

Engineers stress-test the newest models with games and physics sims. I test them against P&L problems: real companies, real customers, real money. Each of these is a running experiment I build, operate, and document.

MIGA Bakery — live website
Live · miga-atl.com

MIGA Bakery

An artisan bakery's entire commercial engine, built end-to-end: the website, payments, SMS, and CRM — plus an agent-supported marketing engine and an operational backbone wired through Airtable. A one-person ERP, run in production.

Cloudflare Workers · Hono · D1 · Stripe · Twilio · Airtable
Visit ↗
Vivoccidente — live map portal
Live · building in public

Vivoccidente

A map-first property portal for the parcelaciones of Antioquia's western highlands — browse the communities around Sopetrán and San Jerónimo on an interactive map: the place first, the property second. The current experiment on the bench, growing week by week.

Leaflet · OpenStreetMap · self-hosted
Visit ↗
Built

EvidenceVault

WhatsApp conversation reconstruction for legal evidence: chain-of-custody logging, SHA-256 verification, audio transcription, and court-ready PDF exports. Built because a real case needed it.

Python · SQLite · Whisper · FFmpeg
Running · next up

The career database

A provenance-tracked database of my own career: every claim sourced, contested facts flagged, guardrails against embellishment. It already writes my CVs. Next: an agent you can interrogate instead of reading one — ask my career anything, grounded in the record.

SQLite · Python · LLM with provenance guardrails

03 — Writing

Lab notes.

Jul 2026 Applied AI

Where are all the business people?

Engineers are stress-testing the newest models with games, physics sims, and benchmarks. The people who own P&L problems mostly aren't experimenting at all — and the most interesting tests are theirs to run.

→

Next up: the full build story of the cash-application agent.

04 — About

The seam is
the job.

My career happens at the seam between what a technology can do and what an organization will actually adopt. Post-merger integration and AI deployment are the same problem wearing different clothes.

I spent nine years at Accenture making two systems, two teams, and two ways of operating become one that functions on Monday morning — including designing and building a 17-country shared-services center from scratch. Then I co-founded an EU-backed AI startup and shipped multi-agent LLM systems years before enterprises began piloting them — production experience with what these models actually do, where they break, and the distance between a demo and a system that survives contact with real operations.

Most recently I returned to commercialize the shared-services center I once designed — closing its first Fortune 500 US reference with an AI-accelerated diligence process — and built a production LLM agent that applied ~$100M of cash in its first two days.

Now I work as an independent applied-AI operator, and I run the lab: real businesses used as test benches for what AI models can do when you point them at actual P&L problems.

05 — Contact

Put an agent
to work.

Two conversations are welcome here: you want agents deployed inside a live organization — scope to production, adoption included — or you're building a team that does this and want someone who already has. Either way, write.

Email [email protected]
LinkedIn linkedin.com/in/laurawarjanga
Location Atlanta, GA · working across US, LatAm & EU