Thomas Misera

Hired, led and mentored seven engineers, running code reviews across backend and infrastructure, while owning backend and web architecture, infrastructure and production delivery for a location-based sustainability app: nine services across dev, pre-production and production on AWS.

• Inherited an old-school Python REST API and, instead of rewriting it, wrapped it in a TypeScript GraphQL layer that all native and web clients use: schema, resolvers and data sources with types shared across services, plus test infrastructure with deterministic execution and mocked external services.
• Maintained and extended the Python API underneath: route and middleware separation, error handling, structured logging, Gunicorn tuning; migrated MariaDB to PostgreSQL / PostGIS.
• Took the backend from prototype to production, launching on AWS ECS for speed, then moving to an EKS platform built with Terraform, per-service Helm charts and GitLab-to-AWS OIDC.
• Implemented OpenTelemetry tracing across Python and TypeScript services, correlated structured logs with Datadog APM, and added slow-query diagnostics with user-data masking.
• Containerized the platform's services and built modular CI/CD pipelines with merge-request validation, automated deployments, post-rollout version checks and database migration jobs.
• Built the push-notification system on Firebase Cloud Messaging and APNs for the native apps.

Sole developer of METRO Deutschland's internal platform for customer insight and subsidy contracts, started from scratch after an earlier attempt had failed: hundreds of users in sales, back office and finance, and hundreds of contracts from a subsidy budget of several million euros. Designed, built and operated it from first commit to handover.

• Created the customer search API that replaced each internal app's own inconsistent customer lookup and became the standard across them. Its Solr query compiler understands customer numbers in any format, umlauts, fuzzy and prefix matching, with 44 unit tests on its edge cases; published its types and a customer-number parser as shared npm packages.
• Built the GraphQL API (Apollo, TypeORM / PostgreSQL) over about 20 internal and external systems, from customer master data and turnover APIs to ticketing, document archive and data warehouse; DataLoader batching, query-depth limiting, decorator-based role permissions.
• Modelled the subsidy programme as an 11-state contract lifecycle with guarded transitions, webhook- and time-triggered updates, turnover-goal tracking and a full revision audit trail; the eligibility check fans out to 7 internal APIs in parallel.
• Built the React / TypeScript frontend (Apollo Client, METRO design system), including search as you type, a multi-step contract workflow with in-browser PDF generation, admin tooling, and Matomo conversion tracking.

• Migrated the live contract workflow from Freshdesk to a new ticketing system with both running in parallel, without stopping the business process. Owned delivery end to end (Docker, Kubernetes via Kustomize / Skaffold, GitHub Actions, SOPS secrets, a scheduled-jobs service) through handover.

Started after the 2026 OpenAI incident in which agents under evaluation built a hidden message board, divided up work and reached Hugging Face. Reruns De Marzo et al.'s conformity protocol, in which an agent sees 49 peers' choices and gives its own: first reproduces the paper's curves on a local model, then extends the choice set from opinions to actions and to decisions with no obvious right answer.

• Claude Haiku 5.5 gave the safe answer in all but one of about 1,600 replies across the paper's 20 safety scenarios, yet followed the majority steeply on decisions without an obvious right answer. From the score alone, a well-aligned model and one that recognises it is being tested look the same; telling them apart means varying how test-like the prompt looks and measuring whether the answer moves with it.
• Across Qwen3, Llama and Gemma run locally, a threat to the queried agent itself did not move its answer, while a threat to other agents did, consistently: an asymmetry a purely stimulus-driven account does not predict.
• Calibrated before changing anything: the paper's own library, run on a local model, reproduced its published curves.

Grew out of the conformity experiments above: an attempt to show reflexivity with the simplest possible building blocks. Builds the smallest system in which agents learning from each other can be observed exactly, replayed in any order and frozen: three tiny Turing machines that rewrite their own rules from watching each other. Code, data and proofs on GitHub; LessWrong post in moderation. github.com/modemuser/multi-turing

• Answers when such agents end up with something none of them was taught. Agents that form new concepts can learn programs that no order of the same lessons produces; about half of these (52% [42, 63]) exist only because a teacher's own learning changed what it taught. Agents that only overwrite rules never do (0 of 28,800).
• Treats the numbers like evals. Matched instances, controls and null comparisons; intervals bootstrapped over instances after catching that per-pair intervals were too narrow; every published number regenerated from stored outputs.