Projects
Document intelligence for private markets
Fund managers upload fund documents; investors get structured, comparable fund records. Every value traces back to the exact spot on the page it came from, and a person approves it before it counts.
+ − Details
- An extraction pipeline: parse, embed, rank, rerank, then one LLM prompt per field category. Each quote the model returns is checked against the source text
- A second, independent read of the whole document, merged with the first field by field
- A self-hosted retrieval stack (document parser, embedder, reranker) on GPUs that shut down when idle; the LLM is chosen per job
- Provenance to a bounding box on the page, and every prompt, response and token count stored
- Automatic scoring of each extraction against a verified version of the same fund
- Python
- FastAPI
- PostgreSQL / pgvector
- Next.js
- Docling
- Modal
- LLMs
- RAG
- Document AI
- Evaluation harnesses
- Data provenance
Vision Suite
A computer-vision platform for parcel sortation centres. Cameras in the scanning tunnel, GPU workstations on site, one risk score per parcel. Its first product finds counterfeit and illicit goods; others run on the same suite.
+ − Details
- Each parcel photographed from 5–6 sides and read in under half a second: label text by OCR, hazard markings, box shape and visual tells
- A risk score per parcel over a JSON API; parcels above the threshold are diverted for inspection, with an evidence package for the brand to review
- Air-gapped GPU workstations inside the facility, or the cloud where allowed: images never have to leave the building
- A database of 100,000+ known infringers behind the scoring, and random search to catch the unknown
- Seven modules on one suite: illicit and counterfeit parcels, label fraud, limited-quantity and ADR dangerous goods, non-conveyables (by vision and by OCR), no-reads and material classification
- Adapted for Customs and border protection: drugs, firearms and other prohibited items
- Certification, security-governance and due-diligence processes, aligned with GDPR, ISO 27001 and SOC 2
- PyTorch / TorchScript
- YOLO
- OpenCV
- EasyOCR
- CoAtNet
- REST / JSON
- Computer vision
- OCR
- Docker
- On-site deployment
Illicit-trade detection platform
Parcels are photographed as they pass through sorting centers. The platform flags likely counterfeits, weapons and drugs against the brands' reference data, and brands and authorities follow each case through to seizure.
+ − Details
- An AI framework for high-speed logistics: computer vision on every parcel image, in real time
- Hybrid cloud and on-site systems, highly secured, processing 3B+ parcel images
- The platform for brands and authorities: risk analysis per parcel, reference databases, and each case tracked from detection to seizure
- A package of the core technology for customs organizations
- Certification, security governance and due-diligence processes
- PyTorch
- OpenCV
- YOLO
- EasyOCR
- CoAtNet
- Computer vision
- Real-time processing
- Hybrid cloud
- On-premise
- Security & compliance
Bin-packing algorithm and 3D engine
A proprietary bin-packing algorithm that plans how boxes fill pallets and containers, and an in-house 3D engine that turns each plan into step-by-step loading instructions.
+ − Details
- The packing algorithm: box sizes, weights and orientation rules in, pallet and container loads out
- A 3D render engine: every pallet and every loading step shown from the left, the right and the top
- Packing instructions and a CO₂ certificate per order, inside one SaaS suite with APIs
- Cloud-native infrastructure with fully automated CI/CD and QA pipelines
- A remote-first team across Latin America, Europe and MENA, 3× in under a year
- C#
- .NET / Blazor
- Unity
- Three.js
- TypeScript
- WebAssembly
- Optimization algorithms
- 3D rendering
- SaaS
- APIs
AI automations for a company builder
A set of AI automations for the company's daily work. The most valuable one validates new venture ideas automatically, through Beam's proprietary framework.
+ − Details
- Automatic idea validation: each new venture idea is assessed step by step against Beam's proprietary framework
- A shared knowledge base (RAG) that people query and add to from Slack, Telegram, WhatsApp or the web
- HR: candidate processing, from application to a structured profile
- Marketing: content drafted from the company's own knowledge
- n8n
- LLMs
- RAG
- AI agents
- Slack
- Telegram
- WhatsApp
- Workflow automation
Parcel-flow forecasting and load balancing
My longest project at Beam, over several months. Predict parcel volumes across a courier network, then balance the load with an AI algorithm and incentives for end customers, to raise capacity use and cut emissions.
+ − Details
- Predictive models of parcel volumes across a courier, express and parcel (CEP) network
- An AI algorithm that balances the load on the network, driven by those forecasts
- Incentives for end customers to shift deliveries to where capacity is free
- Capacity-use and sustainability metrics to measure the effect
- Python
- Time-series forecasting
- Machine learning
- Optimization
- Load balancing
- Data pipelines
- Sustainability metrics
Venture building
As Entrepreneur in Residence at a VC and company builder backed by BEUMER Group, I found new business ideas in logistics and validated them: the problem, the value, the willingness to pay, each step run through Beam's framework.
+ − Details
- Idea validation through Beam's framework: problem, value, willingness to pay
- Local micro-fulfilment: reusing shop space left empty in the pandemic for local storage and last-mile delivery, to raise margins and cut emissions
- Supply-chain compliance: AI-assisted decisions and risk management to meet supply-chain laws such as the German Lieferkettengesetz
- Venture design
- Customer discovery
- Value proposition
- Willingness to pay
- Business modelling
- Applied AI
PAR ranking platform
The data platform behind a gender-equity ranking of organizations in Colombia, Mexico and Peru.
+ − Details
- The first data systems and architecture of the PAR ranking
- A service-oriented, decoupled re-architecture for scale and reliability
- A data warehouse that raised the company's AI maturity from 12 % to 22 % (M.Sc. thesis case)
- Scrum, OKR and delivery metrics across a remote team in 5+ countries
- Service-oriented architecture
- Data warehouse
- ETL pipelines
- KPI dashboards
- CI/CD
- Cloud
Axioma framework
One CMS framework under a studio's whole output, so each project took less time than the last.
+ − Details
- A proprietary CMS framework used across the studio's projects
- A tenant-aware SaaS platform, a data warehouse with ETL, and an in-house e-commerce engine
- PHP / Laravel
- CMS framework
- Multi-tenant SaaS
- E-commerce
- Multilingual sites
- Web design
River-flow forecast
Forecasting river flow in Kazakhstan, with a researcher of the Potsdam Institute for Climate Impact Research.
+ − Details
- Python
- Time-series forecasting
- Machine learning
- Statistical modelling
- Scientific data
In 2012 I founded Axioma Estudio SAS (axiomaestudio.com | axioma.ai), a custom design & software development agency. I personally built 120+ projects during the first 6 years, and I wrote our proprietary framework so each project took less time than the last. The development process went from several weeks to days.
