main abstraction

Lisbon · remote across the EU

Your batch jobs take longer every month and one engineer is the only person who can fix them.

I am Gonçalo Traça. Two years at Accenture, data engineer on national-scale telecom pipelines. I rebuilt the processing path and cut end-to-end runtime by about 85%. Six years of cloud and data engineering in total, and an MSc in Big Data. I measure what your pipelines cost and how long they take, fix one bottleneck, and measure it again the same way.

GCP · BigQuery · AWS · Postgres · Python · SQL

what we do

Four things. Pipelines and cloud is the one I have been paid to do for six years. For the other three I say plainly what I have shipped and what I have only measured.

Data pipelines and cloud

Ingestion, scheduling, reprocessing, backfills, retries, alerting. GCP, BigQuery, AWS, Postgres, Python, SQL. At Accenture this was national-scale telecom data. I rebuilt the processing path and cut end-to-end runtime by about 85%.

Retrieval and evals

Retrieval and agents over your own documents, scored against an eval set built from your own examples. I have measured retrieval over three corpora: 7,591 job postings, 150,881 EU procurement notices in 23 languages, 270,000 legal paragraphs. On two of them the hybrid approach lost to plain BM25 keyword search. I published that. I have not been paid to do this work yet. The measurements are public and you can check every number.

On-chain systems

Solidity and TypeScript contract work for DeFi protocols from 2022 to 2024, including Poolshark. Blockchain integrations and AWS data pipelines at Celfocus. ScorePlay runs onchain on Base: 120,000 users, 1.2 million predictions. I founded it, raised a $150k pre-seed, and led the team of seven that shipped it.

Twelve days, fixed scope

Twelve working days, one fixed price, no change requests. Days 1 to 3 I instrument and measure what your pipelines cost, how long they take, and how often they fail. On day 3 we pick one bottleneck together. Days 4 to 10 I fix it and ship it. Days 11 and 12 I re-measure the same way and write the handover. Tell me what you are running and I will send the number the same day.

Design · Build · Review

selected work

Two projects documented in full, from the requirement that shaped them to the numbers they were measured against and the limits they still carry. Different domains, one method.

Example project · applied AI

A trust contract for generated reports

A product that turns public signals into a structured verdict, where each claim carries a confidence label derived from the reliability of its source rather than from the model’s own assessment of itself.

Self-reported confidence is generated text like everything else, and it tracks how fluent a claim sounds rather than whether it was observed. Labels are computed in code, per claim rather than per report. The evaluation surface is split along determinism: a deterministic contract suite runs on every commit at zero cost, and the expensive model-facing suite runs nightly, because the failures they catch do not overlap.

2: evaluation layers, split by cost
46: deterministic cases per commit
<1s: contract suite runtime

Read the technical report

Example project · applied AI

Retrieval over a closed document collection

A question-answering system over a fixed corpus where every claim is traceable to the passage it came from. Hybrid retrieval with reranking, a bounded tool-calling agent, and three verification checks that run in code after generation.

Built around one requirement: an answer has to be checkable. Citations must resolve to passages retrieved that turn, quotations are string-matched against the sources, and each sentence is audited against its evidence. Measured on three evaluation sets of increasing difficulty, one never tuned against, with differences decided by paired bootstrap rather than by score movement.

52: evaluation cases
3: verification checks in code
2: inherited defaults corrected

Read the technical report

Example project · blockchain infrastructure

Access-control libraries for a zkVM execution layer

Two Rust libraries for an SVM-style execution layer: an admin authority and a freeze authority (circuit breaker). Foundational, security-relevant code on a platform where programs compile to RISC Zero zkVM guests.

We built the libraries, then proved them: compiled to zkVM guests, deployed on a local sequencer, every privileged path and every rejection exercised on-chain and checked against the execution log. 60 host tests green, CI green, dual-licensed MIT and Apache 2.0. Working that close to the tooling also surfaced a bug in the framework’s own CLI, which we traced, patched, and submitted upstream.

2: libraries delivered
60: host tests passing
10: instructions verified on-chain

Read the technical report

Design · Build · Review, on one project

Questions

What does Main Abstraction actually do?: Data pipelines and cloud work, mostly. Ingestion, scheduling, reprocessing, backfills, alerting, and making slow things faster. Retrieval and eval work on top of that when a team wants AI over their own documents and needs a number rather than a demo.
Who does the work?: I do. Main Abstraction is one person, Gonçalo Traça. If a job needs more hands than I have, I say so before we start rather than quietly subcontracting it.
How do you usually work?: Twelve working days, one fixed price, no change requests. Days 1 to 3 I instrument and measure. On day 3 we pick one bottleneck together. Days 4 to 10 I fix it and ship it. Days 11 and 12 I re-measure the same way and write the handover. Longer engagements exist but this is how they start.
How is pricing handled?: Fixed price per engagement, agreed before anything starts. Tell me what you are running and I will send a number the same day. Day rate if you prefer it, though fixed scope is usually the better buy for both sides.
Is there a minimum commitment?: No retainer, no minimum term. The twelve-day engagement is designed to end cleanly. Most people who want more ask afterwards.
Have you done this for other clients?: Not through Main Abstraction, which is a new company. The six years of pipeline work behind it were done on payroll, at Accenture and Celfocus, on national-scale telecom data. The retrieval and eval work is my own, published, with every measurement public and checkable. I would rather tell you that than have you find out.
Where are you based, and do you work remotely?: Lisbon. Remote across the EU, and on site for a kick-off if you want one and travel is covered. Portuguese, English, and workable Spanish.
Who owns what you build?: You do. Code, documentation, and the measurements. Assigned on final payment.
What happens after delivery?: You get a written handover: what changed, why, what I chose not to do, and what breaks next. Written so another engineer can pick it up. I would rather be recommended than depended on.
How do we contract, and what about VAT?: Main Abstraction is a Portuguese company with an EU VAT number. Business to business inside the EU is reverse charge, so there is no permanent establishment created for you and no local payroll involved.

about

Main Abstraction is a technical practice. We take a few projects at a time and build them end to end, from first call to production.

The work runs from retrieval systems that answer from a client’s own documents, to automation that removes manual steps between tools, to security-relevant libraries deployed and verified on-chain, across sectors from energy and telecoms to blockchain infrastructure.

Est. 2024

get in touch

Tell us what is eating your team’s time. We will tell you if we can automate it.

gon@mainabstraction.com

Questions

What does Main Abstraction actually do?

Data pipelines and cloud work, mostly. Ingestion, scheduling, reprocessing, backfills, alerting, and making slow things faster. Retrieval and eval work on top of that when a team wants AI over their own documents and needs a number rather than a demo.

Who does the work?

I do. Main Abstraction is one person, Gonçalo Traça. If a job needs more hands than I have, I say so before we start rather than quietly subcontracting it.

How do you usually work?

Twelve working days, one fixed price, no change requests. Days 1 to 3 I instrument and measure. On day 3 we pick one bottleneck together. Days 4 to 10 I fix it and ship it. Days 11 and 12 I re-measure the same way and write the handover. Longer engagements exist but this is how they start.

How is pricing handled?

Fixed price per engagement, agreed before anything starts. Tell me what you are running and I will send a number the same day. Day rate if you prefer it, though fixed scope is usually the better buy for both sides.

Is there a minimum commitment?

No retainer, no minimum term. The twelve-day engagement is designed to end cleanly. Most people who want more ask afterwards.

Have you done this for other clients?

Not through Main Abstraction, which is a new company. The six years of pipeline work behind it were done on payroll, at Accenture and Celfocus, on national-scale telecom data. The retrieval and eval work is my own, published, with every measurement public and checkable. I would rather tell you that than have you find out.

Where are you based, and do you work remotely?

Lisbon. Remote across the EU, and on site for a kick-off if you want one and travel is covered. Portuguese, English, and workable Spanish.

Who owns what you build?

You do. Code, documentation, and the measurements. Assigned on final payment.

What happens after delivery?

You get a written handover: what changed, why, what I chose not to do, and what breaks next. Written so another engineer can pick it up. I would rather be recommended than depended on.

How do we contract, and what about VAT?

Main Abstraction is a Portuguese company with an EU VAT number. Business to business inside the EU is reverse charge, so there is no permanent establishment created for you and no local payroll involved.