Hugo Sanches

Senior RPA developer. I build the software robots behind banking operations, and connect them to generative AI.

Sintra, Portugal. Currently at act digital.

See selected work Get in touch

Move your cursor to stir the water.

Portrait of Hugo Sanches

From tsunamis to software robots

I started out writing algorithms to detect tsunamis from space, combining radar altimetry with reflectometry. Then came electricity markets, forecasting prices and consumption with machine learning. Since 2019 it has been software robots: scaling an RPA programme from proof of concept to more than 50 bots at Accenture, keeping a bank's 24/7 automation estate stable at Millennium bcp, and now building RPA and generative AI solutions for a major international financial group at act digital. Different water, same job: find the signal, then make it act.

<1%
error rate in production
50+
bots running at once, grown from proof of concept
4
people in the production control team I grew from one
24/7
automation estate kept running

Journey

  1. 2026now

    Senior RPA Developer act digital

    RPA and generative AI solutions for a major international player in the financial sector, bringing years of banking automation to a new team.

  2. 20222026

    Senior RPA Developer and Controller Millennium bcp

    Joined to scale the bank's large RPA programme and strengthen its governance and security, in a complex 24/7 environment.

    • Led and scaled the production control team to a production error rate under 1%.
    • Pioneered generative AI in the automation stack.
    • Built monitoring and analytics tools for faster fixes and proactive maintenance.
    • Hardened the robot infrastructure with new security, resilience and redundancy measures.
    • Saved thousands of manual work hours a year while meeting strict SLAs.

  3. 20192022

    Senior Analyst, Business and Integration Architecture Accenture Portugal

    Associate Software Analyst, then Software Analyst, then Senior Analyst

    Joined an early-stage RPA initiative with only proof-of-concept bots live, as its one-person production control.

    • Built and led the production control team, from 1 to 4 people.
    • Helped scale the programme to more than 50 bots running at once.
    • Introduced OCR (ABBYY) and NLP solutions.
    • Supported several development teams with release management and incident resolution.

  4. 20162018

    Developer, then IT Officer Elusa, utility services start-up

    Forecast electricity market prices and client consumption with machine learning, processed large volumes of data, automated company processes and built the web presence.

  5. 20162017

    Research partnership University of Lisbon

    New algorithms combining reflectometry and radar altimetry to detect tsunamis from space.

  6. 2015

    Developer Magicoach

    Software to analyse data on professional football players.

  7. 20102019

    Meteorology, Oceanography and Geophysics University of Lisbon

    Bachelor's degree (2010 to 2014), followed by master's studies in Geophysics, specialising in oceanography.

Selected work

Six problems from banking automation, and how I approached them. Client details stay general; the engineering is real.

Keep scrolling. The work moves sideways.

Teaching robots to speak mainframe

Core banking still runs behind IBM 3270 green screens, and the robots needed a dependable way in.

I built a C# automation library for IBM Personal Communications, used from Blue Prism: 25 low-level primitives behind 17 published actions.

  • Two layers: primitives in shared code, orchestration in actions.
  • Idle-session detection, protection against re-login loops, launch timeouts.
  • Hardened through a structured code review, fix by fix.

C#, IBM PCOMM, 3270, Blue Prism

 CORE BANKING             SESSION A
 ----------------------------------
 TRANSACTION ==> ____
 ACCOUNT     ==> ______________

 STATUS      ==>                   



 PF3 EXIT  PF7 UP  PF8 DOWN  ENTER

    Moving email automation to Microsoft Graph

    Driving Outlook through COM on robot machines is fragile: profiles, MAPI sessions, DNS. Graph takes the desktop client out of the loop.

    I built Graph API actions for Blue Prism with app-only authentication (MSAL.NET and certificates), OData filtering, and Microsoft Information Protection sensitivity labels on send, reply and forward.

    • Resolved Exchange Online application access policy errors with the Exchange admins.
    • Worked around OData limits on multi-sender queries.
    • Kept the legacy COM object stable meanwhile, with full retry and recovery.

    Microsoft Graph, MSAL.NET, C#, VB.NET, MIP labels

    
                
    Sensitivity label Confidential
    
              
    
            
    
          
    

          
          
    
            
    
              
    Fewer tokens per robot conversation
    
              
    When robots call a language model thousands of times a day, all of JSON's punctuation becomes a cost line.
    
              
    I implemented TOON (Token-Oriented Object Notation) as a C# serializer for Blue Prism payloads sent to LLMs, and validated it on an invoice-validation prompt.
    
              
      
                
    • Same data, far fewer characters than JSON for tabular payloads.
      • 
                
    • Grew it into a summer internship I designed and mentored: benchmarking output formats for the team's LLM agents.
      • 
              
    
              
    C#, LLMs, prompt design, cost optimisation
    
            
    
            
    
              
    
                
    
                  
    JSON 0 characters
    
                  
    
                
    
                
    
                  
    TOON 0 characters
    
                  
    
                
    
                
                
    0% fewer characters than minified JSON
    
                
    
                  
                  
                
    
              
    
            
    
          
    

          
          
    
            
    
              
    Reading thousands of credit contracts a month
    
              
    Scanned credit contracts were going to a commercial LLM API for field extraction, at a cost that grows with every page.
    
              
    I led the evaluation of self-hosted vision-language models to bring extraction in-house, from DONUT and Qwen2.5-VL to PaddleOCR-VL and GLM-OCR, targeting GPU endpoints on Azure ML.
    
              
      
                
    • Designed a hybrid pipeline: template classifier, extraction, signature verification and tamper detection.
      • 
                
    • GDPR requirements and a human review step built into the design.
      • 
              
    
              
    Python, vision-language models, OCR, Azure ML
    
            
    
            
    
              
    
                
                
    
                  
    Contract
    CR-2026-0481
    
                  
    Amount
    €25,000
    
                  
    Term
    84 months
    
                  
    Signature
    Verified
    
                
    
              
    
            
    
          
    

          
          
    
            
    
              
    Robots that answer in seconds, not hours
    
              
    Blue Prism is built for batch work. A personal-loan decision needs an answer in under 30 seconds.
    
              
    I wrote the proposal and built a working demo for near-real-time processing: work triggered through the Blue Prism API, brokered by RabbitMQ, on high-availability work queues.
    
              
    Blue Prism API, RabbitMQ, solution architecture
    
            
    
            
    
              
    
                
                
    0.0 s
    
                
    Simulated. Target: under 30 seconds per request.
    
              
    
            
    
          
    

          
          
    
            
    
              
    Letting AI read the robots' source
    
              
    Blue Prism processes live in XML that is hard to review and harder to document.
    
              
    I'm building LLM agents that analyse Blue Prism process and object XML, and turn it into Draw.io diagrams using stage-to-shape mapping rules.
    
              
    LLM agents, prompt engineering, Blue Prism XML, Draw.io
    
            
    
            
    
              
    
                
                
                  
                  
                  
                  
                  
                  
                  
                  
                  
                
              
    
            
    
          
    

        
    
        
      
    
    
    

    
    
    
      
    
        
    Beyond the code
    
        
    
          
    
            
    Building teams from one
    
            
    At Accenture I started as the only production controller and grew it into a team of four. At Millennium bcp I led and scaled production control for a 24/7 estate.
    
          
    
          
    
            
    Mentoring
    
            
    I designed and ran a summer internship end to end: the proposal, onboarding, kickoff and plan, in English and Portuguese, mentoring an intern on LLM token optimisation.
    
          
    
          
    
            
    Review culture
    
            
    I introduced an RPA-specific code review checklist, and put my own libraries through it first.
    
          
    
          
    
            
    No single points of failure
    
            
    I run formal knowledge transfer on critical processes, so no automation depends on one person.
    
          
    
        
    
      
    
    
    

    
    
    
      
    
        
    Toolbox
    
        
    
          
    Automation
    Blue Prism, IBM PCOMM and 3270, Microsoft Graph, Outlook and Excel COM, n8n
    
          
    AI
    LLM integration, RAG, prompt engineering, intelligent document processing, OCR with ABBYY, Gemini and vision-language models
    
          
    Code
    C#, VB.NET, Python, Selenium, VBA, Django
    
          
    Platforms
    Azure ML, REST APIs, RabbitMQ, MSAL and OAuth
    
          
    Practice
    Production control, release management, Agile, design thinking
    
          
    Languages
    Portuguese (native), English (fluent)
    
          
    Certification
    n8n Course, Level 1
    
        
    
      
    
    
    

    
    
    
      
    
        
    Let's talk automation.
    
        
    If you're building automation or generative AI in a regulated business, I'd like to hear about it.
    
        
        
    
          
          LinkedIn profile