Adrian Matias

Senior engineer on an LLM-as-judge product that scores software vendors. Built the human ground-truth lane alone — domain-expert template parser, vendor matching, band placement, tier-weighted averaging and roll-up — the lane that reconciles human evaluator scores against model scores and makes the judge's accuracy checkable against ground truth.

• Contributed to every component of the product architecture in a team of three senior AI engineers, a PM and the associate product owner.
• Introduced and owned schema-driven, domain-driven and spec-driven development: business requirements kept as a contract artifact that agentic development builds against.
• Owned the typed output contract and the correctness gate set — type baselines per file and rule, a complexity ratchet, mutation tests, required status checks — and designed and managed the end-to-end validation workflows behind them.
• Created and owned full epics end to end, from business requirement to shipped increment.
• Ran agentic development at production scale as a measured practice: cost and turn discipline tuned per model tier.

Install prediction for mobile user acquisition: cold-start, transformers, user embeddings, transfer learning, MLOps, production pipelines, parity resolution, experimentation, monitoring. Progressed toward technical leadership on ownership and strategic perspective.

• Unified modeling across targets, advertisers, and platforms via attention/transformer architecture, increasing end KPI.
• Design and deploy cold start for advertiser onboarding in two angles: transfer learning through an affinity tower trained on wider set of interactions and cold advertiser masking validated on held-out cold entities. New advertisers bid from day one from a unified model, with eCPI better than curated whitelists for inventory and demand segments. Custom PyTorch
• Resolved parity mismatches and calibration gaps: traced discrepancies in internal transformations until models received identical inputs offline and online and their outputs matched through a complex inference server; productionized parity workflows with unit tests in Databricks.
• Designed and proposed the user embedding system built from users' interactions with apps; realized with transfer learning at performance parity and orders-of-magnitude cost savings.

End-to-end ML for international clients: scope review, stakeholder alignment, model productivization.

• Led rapid turnaround of underperforming client ML system (utility fraud detection & ranking): surgical plan, restored accuracy/efficiency, recovered client satisfaction.
• Surfaced critical issues in data ingestion and feature-engineering pipelines.
• Real-time credit card fraud classifier (classical ML + deep VAE). Open: github.com/adrianmatias/ml-sandbox/tree/master/tx_class

Ad ranking and reinforcement learning for in-app game advertising; DSP performance via large-scale experimentation.

• Bayesian Bandit RL system with policy evaluation tools; tailored online/offline evaluation frameworks.
• Two-Towers neural ranking from recommender-systems patterns.
• Post-mortems on adversarial DSP dynamics and model feedback loops.

Text analytics, NLU, and semantic search on massive financial news data; MLOps and reproducibility.

• Few-shot search personalization: simplified architecture, −30% error.
• Active learning PoC for learn-to-rank: projected +300% annotation efficiency.
• Long-text fingerprinting for duplicate content detection.