Telmo da Silva

Technical training for industrial manufacturing teams, in a blended model: e-learning first, then Q&A sessions with up to 50 trainees for common questions, then hands-on practice split by track (frontend, backend, automation, administration) with a dedicated trainer per track.

• Close to 90 training runs and 1,800+ trainees over four years (around 23 a year, cohorts of 20 to 30), with an approximate 85% pass rate.
• Trainer for the Modeling track, hands-on configuration of the manufacturing execution system (MES), certified as an Associate Modeler.
• Automated training administration in Python: querying the participant database by edition, joining it with final exam results, and extracting the certified list, replacing manual spreadsheet filtering.

My own studio, always built alongside a full-time job. It is where I develop and run Moto Spotters and Scryve in production, and where the practice I teach comes from.

• Multi-agent adversarial auditing before every release: a second agent is instructed to refute each finding from the first, not confirm it, and defaults to discarding it on any failure. It surfaced real vulnerabilities I had missed reviewing alone, and nothing counts as fixed without a test that reproduces the attack.
• Systems where a mistake stops being recoverable once published: protocols and payment infrastructure across several networks (Cardano, Nervos CKB, Bitcoin, EVM), in Rust.

Community platform for motorcyclists in Portugal, live in production (Next.js, PostgreSQL).

• LLM-based content moderation across 10 categories, called from 73 places across 29 files: three severity-weighted cut-offs, not one, where only the highest always blocks. Calibrated in shadow mode, scoring without blocking, against real traffic before enforcing.
• Ride passport with organizer-verified stamps, motorcycle-club management and a peer-to-peer marketplace.

Long-form publishing platform, live in production: an AI writing assistant and a custom micropayments system.

• Plagiarism detection using per-paragraph embeddings and cosine similarity at a 0.75 threshold. Per paragraph rather than per article, because one copied paragraph in ten is diluted by the nine original ones and disappears from the average.
• Custodial micropayment rails on Nervos CKB: per-user deposit wallets swept to a treasury, and queued withdrawals settled as one batched on-chain transaction rather than one per recipient, so each additional payout costs a single output instead of a full transaction fee.
• Crash-safe payout queue: the transaction hash is computed and written to the database before submission, since CKB hashes the raw transaction excluding witnesses. A crash between submitting and recording cannot double-pay; orphaned claims return to the queue and a rejected batch restores every balance.

Induction and product training for new hires in a customer experience operation, assessing each person's readiness before they went live.

• Around 170 cohorts and 1,600+ employees trained over seven years: a new cohort every two weeks, 10 to 15 people each.
• Four-week follow-up per cohort in live production, often while already running the next one: anyone who showed a serious gap was pulled, coached, and only returned to the floor afterward.
• As Quality Team Leader, operations management with Quality Coaches, Trainers and SMEs reporting to me, across teams of up to 100 people.

Progressed from agent to SME to Quality Coach: grading contacts against four indicators (QA score, critical error rate, compliance, first-contact resolution), root-cause analysis by error category, client calibrations, and one-on-one coaching.

270+ chapters, free, no account required. 4 courses on Python for QA and 4 on AI Engineering (production AI APIs, RAG, evaluation & observability, agents), 76 chapters with code. I built the assessment layer myself: Python runs in the learner's own browser, CPython compiled to WebAssembly inside a web worker, coding challenges are graded against 18 hidden tests, and the certificate is only issued once they pass. Each course tested end to end before publishing.

Rules stored as data, each citing the published document it comes from, so any rule can be contested against its source rather than against my taste. Exits with code 1 when it finds anything, so it can live in a commit hook or a CI step: a validator that reports but does not stop is a report, not a gate. 25 tests, and the build fails if the linter stops catching a deliberately broken example, which means the gate is tested too.