Vasco F.-Senior AI Engineer

Databricks Architecture & AI Data Products [Databricks Apps · Unity Catalog · Delta Lake · Lakebase · Asset Bundles · Azure DevOps · MCP]

• Designed a layered Databricks financial data architecture: Apps/FastAPI for serving, modular Python calculation engines, SQL/Delta for governed analytics, and Unity Catalog Volumes for versioned workbook outputs.
• Integrated Lakebase PostgreSQL operational data with SQL warehouse analytics, keeping transactional and analytical access separate through App service-principal identity, short-lived credentials and controlled, read-only data interfaces.
• Defined DEV/TST/PRD deployment architecture using Databricks Asset Bundles and automatic Azure DevOps CI/CD, with reusable environment configurations, OAuth M2M identities and automated application/resource lifecycle management.
• Designed access boundaries across Apps, SQL warehouses, Unity Catalog tables/functions/Volumes and secret scopes; separated deployment, application and SQL identities through declarative resource permissions and environment-specific authentication.
• Built serverless Spark pipelines for geospatial reference data and a traceable storage architecture combining immutable Volume artifacts, revision lineage and append-only Delta history with durable audit-event delivery.
• Connected AI Hub agents to governed Databricks financial workflows through typed REST/MCP interfaces; implemented CAPEX/OPEX, commercial, lease and IRR/NPV engines with scenario validation and workbook consistency checks.

EGA & Audit Work-Plan Agentic Workflow [LangGraph · LangChain · Azure OpenAI · Azure Document Intelligence · pgvector · Azure Databricks · Streamlit]

• Architected and delivered a production agentic audit workflow built as a LangGraph supervisor orchestrating specialist workers: a Document Parser Agent (PDF/table extraction via Azure Document Intelligence + Apache Tika), a Control-Mapping Agent (aligning prior-year Entity Governance Assessment assertions with current-year Procure-to-Pay controls), a Delta-Detection Agent (flagging design, operating-effectiveness, and evidence-gap changes), an Evidence-Validation Agent (multi-modal checks over screenshots and workpapers), and a Citation Agent (source-linked footnote generation); compressing a multi-day manual control-comparison review into real-time auditor dialogue with ground-truthed citations. Adopted across Finance & Internal Audit practices at go-live.
• Equipped the agent with first-class audit tool calls: SQL-over-workpaper queries against Azure SQL, structured JSON extraction from PwC work-plan templates, automated control-walkthrough drafting, risk-register lookup, and one-click export into PwC's internal audit platform; the supervisor orchestrates ~15 discrete steps per auditor question with deterministic handoffs and retry/fallback policies.
• Engineered advanced RAG retrieval achieving (>90%) precision/recall in control-delta detection: hierarchical embeddings (document (\rightarrow) section (\rightarrow) chunk), HyDE-style synthetic query augmentation, cross-encoder re-ranking, and multi-turn conversational memory (sliding window + entity memory) over pgvector-backed PostgreSQL with ACID guarantees and sub-100 ms ANN search.
• Designed the full Azure data architecture: Spark-based PDF/table preprocessing on Azure Databricks (containerised, parameterised DAGs), orchestration via Azure Functions + Event Grid triggers, artefact versioning in Azure Blob Storage, CI/CD in Azure DevOps with automated integration tests; surfaced via a Streamlit dashboard with streaming token output, source-link popovers, and markdown rendering.
• Embedded multi-modal controls validation by batch-vectorising control screenshots and evidence images through GPT-4 Vision, cross-checking narrative assertions against visual artefacts inside the same chat window; a novel audit pattern enabling auditors to flag discrepancies in a single conversational turn.
• Hardened for Big-4 enterprise standards: prompt templating (PII redaction, guardrail moderation), token-cost tracking per engagement, LLM-as-judge regression test suite, GDPR-compliant data retention policies, and Power BI telemetry dashboards; establishing a reusable Gen-AI audit blueprint as a firm-wide AI accelerator.

Autonomous Agent for Embedded Software Development [Anthropic Agent SDK · LangGraph · Claude 3.5 Sonnet · GPT-4 · Azure OpenAI · C/C++ · HIL CI/CD]

• Architected and led delivery of a multi-step autonomous coding agent built on Anthropic's Agent SDK, orchestrating a supervisor plus 5 specialist workers (Code Generator, Static Analyser, Test Scaffolder, HIL Runner, Memory Manager) that automates embedded C/C++ firmware feature development end-to-end: from natural-language specification to hardware-validated commit, reducing average feature-development cycle time by ~30 % and freeing senior embedded engineers from repetitive boilerplate.
• Leveraged the Anthropic Agent SDK's native tool-use primitives and long-context streaming so the supervisor could invoke deterministic sub-tools (compilers, linters, HIL runners) as first-class citizens: Claude 3.5 Sonnet handled planning, self-reflection and critique, while GPT-4 via Azure OpenAI served as a parallel reviewer for cross-model disagreement detection before commits were promoted.
• Integrated MISRA-C 2012 and Clang-Tidy as first-class agent tool calls: the Static Analyser worker runs the full checker suite on generated code, parses diagnostic output into structured JSON, and returns violation descriptions to the Code Generator for targeted self-correction; creating a closed-loop that enforces ISO 26262 / IEC 61508 safety standards without human intervention.
• Connected the agent to ZEISS's hardware-in-the-loop (HIL) CI/CD pipeline via a custom gRPC adapter: the HIL Runner worker flashes generated firmware to a simulated sensor board, streams test-execution telemetry (pass/fail counts, timing violations, register states) back into the LangGraph state machine, and triggers corrective re-generation on failure; achieving a self-healing loop that passes HIL validation in ≤ 3 iterations on >90 % of tasks.
• Made regression testing a first-class gate in the agent loop: every candidate patch is automatically run against the existing C/C++ unit-test suite (GoogleTest/Unity) plus the HIL regression harness, with results fed back as structured diagnostics; the agent is explicitly forbidden from opening a pull request on a patch that breaks a previously-passing test, and delta coverage is tracked per feature so a drop in line/branch coverage blocks promotion.
• Designed a dual-memory architecture: a long-term vector store (ZEISS sensor APIs, BSP docs, coding conventions) for retrieval-augmented code generation, and a short-term session KV-cache preserving cross-turn architectural context; reducing hallucinated hardware register references by >85 % vs. a vanilla GPT-4 baseline.

Time Series Forecasting [N-BEATS · Chronos · Azure ML · Prophet]

• Owned the technical strategy and team leadership for ZEISS's enterprise forecasting capability: managed a team of 4 data scientists and 2 ML engineers across financial and operational forecasting domains, conducting weekly architecture reviews, sprint planning, and individual mentoring. Defined ML system architecture (feature store design, retraining triggers, champion/challenger evaluation framework) and presented quarterly roadmaps to divisional VPs.
• Evaluated and integrated foundation time-series models, Amazon Chronos (zero-shot), N-BEATS, N-HiTS, and TFT, against classical baselines (ARIMA, Prophet, ETS), implementing a rigorous walk-forward cross-validation harness with pinball-loss and WAPE metrics; GenAI-based models outperformed classical approaches on 7/9 target KPIs in financial scenario planning.
• Served as solutions architect for ZEISS's inter-departmental data infrastructure: designed Azure-native feature engineering pipelines (Azure ML, ADF, Databricks Delta Lake, DevOps CI/CD) capable of onboarding new business units within 2 weeks via a self-service schema-registration API.

Manufacturing Process Optimisation [Causal Inference · DoWhy · EconML · OPC UA/SCADA]

• Led a first-of-its-kind causal-AI programme at ZEISS, applying constraint-based (PC/FCI), score-based (GES), and linear non-Gaussian (LiNGAM) structure learning algorithms to dense multi-variate SCADA sensor streams, revealing hidden upstream defect pathways invisible to correlation analysis.
• Architected doubly-robust causal estimators (DML, DR-Learner via EconML) to produce unconfounded ATE/CATE estimates of process-parameter interventions on lens-scratch rate and MTF score; providing engineering teams with do-calculus evidence for targeted machine adjustments that reduced defect rate on the target production line.
• Built Bayesian-network what-if simulators from learned DAGs enabling quality-control engineers to run virtual do-interventions (do(X = x)) before physical trials, cutting physical experiment cycles for new process configurations.

Log Analytics Anomaly Detection [Autoencoders · LSTMs · Transformers · MLOps]

• Architected an enterprise-scale log-analytics platform for pharmaceutical intra-ocular lens production: unsupervised autoencoder + isolation forest ensemble for cross-sensor anomaly detection, LSTM/Transformer models for sequential pattern recognition in unstructured log streams, and an NLP log-parsing layer (Drain3 + sentence embeddings) for semantic anomaly clustering; reducing MTTD by 4 % and alert fatigue by 7 % at TBs/day ingestion.
• Delivered an A/B-testable MLOps pipeline on Azure ML for continuous model monitoring (data/concept drift detection via MMD and PSI), automated retraining triggers, SHAP/LIME explainability artefacts per alert, and a model governance registry; establishing organisational best practices for anomaly-detection AI across ZEISS's manufacturing portfolio.

Programmatic Bidding Platform [RTB · LP · RL · Auction Theory · Real-Time Feature Store]

• Designed and owned the full-stack programmatic bidding system serving $150M+/year in recruitment ad spend across 500+ Fortune 500 employers, from real-time bid decisioning (<100 ms RTB latency) to post-campaign measurement. Defined the technical roadmap, led a team of 3 data scientists and 2 engineers, and presented quarterly strategy to the CTO and product leadership.
• Engineered a production bid-price optimiser (PuLP + custom LP extensions) jointly solving budget pacing (fluid pacing curves), CPC landscape optimisation, and inventory quality constraints across all active campaigns simultaneously; achieving a 6 % reduction in Cost Per Application (CPA) at $10M+ weekly spend and improving budget utilisation from 84 % to 97 %.
• Pioneered a first-price auction adaptation strategy in response to the 2019-2021 industry-wide migration from second-price to first-price auctions: designed a bid-shading algorithm using empirical win-rate curves (KDE on historical clearing prices) and a parametric bid-landscape model to estimate optimal shading coefficients per publisher/placement; preventing over-spend of ~$4M/year on first-price inventory.
• Built a Reinforcement Learning bidding agent (Stable Baselines3 PPO) for dynamic intra-day CPC adjustments: state space encodes real-time bid-win-rate, budget-pace ratio, time-of-day, and conversion propensity; reward is -CPA; action space is a continuous bid-multiplier in [0.5, 2.0]. Trained via simulation on historical RTB logs, validated with shadow-mode evaluation before A/B rollout; delivering ~8 % ROAS uplift on targeted campaigns vs. the LP-only baseline.
• Architected a real-time feature store (Redis + Hive) serving <10 ms feature lookups per bid request: pre-computed publisher quality scores, audience propensity scores, and temporal budget-pacing signals updated on 5-minute micro-batch cycles via Spark Streaming.

Incrementality Measurement & Marketing Attribution [Geo Holdout · DiD · Shapley · MMM · MTA]

• Built Radancy's incrementality measurement capability from the ground up, designing geo-holdout experiments (matched market pairs via synthetic control), time-based holdout studies, and ghost-bidding tests to quantify the true causal lift of programmatic campaigns; replacing a last-click attribution model that was overstating ROAS by ~30 % and redirecting ~$18M in annual budget toward highest-incremental-lift channels.
• Engineered heterogeneous treatment effect pipelines (DoubleML, DR-Learner, X-Learner via EconML/DoWhy) producing audience-segment-level CATE estimates with bootstrap confidence intervals; enabling the bidding system to bid higher on segments with high incremental conversion probability, reducing effective wasted spend by ~12 % on average across accounts.
• Implemented a Markov-chain multi-touch attribution (MTA) model (PySpark) computing removal-effect Shapley values across the full conversion path (search → display → programmatic → retargeting); giving clients channel-level true-contribution CPA and hire-rate metrics replacing fractional last-touch heuristics, and directly informing budget-mix decisions for $150M+ in spend.
• Developed a Bayesian Media Mix Model (MMM) with hierarchical priors (PyMC3), saturation curves (Hill function), and adstock decay to separate long-term brand effects from short-term activation; enabling clients to holistically optimise across programmatic, SEM, and organic channels for the first time.

AdWords Copy Generation & Audience Science [NLP · GenAI · Lookalike Modelling · EKS · FastAPI]

• Engineered a production NLP/GenAI ad-copy generation system: domain ontology construction (NER + topic modelling on 2M job descriptions), SBERT/FAISS entity retrieval for grounding, fine-tuned GPT/T5 for constrained generation enforcing brand-voice style guides; deployed as FastAPI microservices on EKS with Prometheus/Grafana monitoring, generating 100k+ brand-aligned ad headlines per day.
• Built look-alike audience models (LightGBM + calibrated probability outputs) on first-party application-conversion signals, expanding addressable reach to high-propensity candidate audiences at 3 × the conversion rate vs. behavioural interest targeting, directly driving incremental qualified applications at lower CPA.

• Designed and owned a production Spark Scala pipeline on Azure HDInsight / Azure Databricks ingesting ~100 GB/day from 10,000+ connected vehicles via Azure Event Hubs / IoT Hub, with custom Scala UDFs for CAN-bus signal decoding and trip-level feature aggregation; curated telemetry landed in Azure Data Lake Storage Gen2 under a bronze/silver/gold medallion layout for downstream analytics and ML workloads.
• Orchestrated end-to-end training and scoring jobs in Azure Data Factory with parameterised Databricks notebook activities and Git-based CI/CD in Azure DevOps; model artefacts were versioned in Azure Blob Storage and served via Azure ML managed endpoints with automated drift-monitoring hooks.
• Built a driver fuel-efficiency ranking model (Scikit-learn) surfaced to fleet managers through Power BI dashboards wired to the Delta Lake gold layer, and led early predictive-maintenance anomaly-detection research using isolation forests on CAN-bus telemetry to flag out-of-distribution sensor patterns before component failure.

• Supported client-facing data analytics engagements: built SQL-based reporting pipelines, automated recurring Excel dashboards in Python, and produced data visualisations (Tableau) for internal audit and advisory teams.
• Gained early exposure to structured data analysis workflows in a Big-4 professional services environment, working across finance and operations datasets for Irish and multinational clients.