Alexander Alyushin

Contributing upstream to ZeroClaw, a Rust AI agent runtime (~31K stars on GitHub, Apache 2.0/MIT). Self-hosted infrastructure: the agent, the data, and the machine all stay with the user.

Sole ML/AI engineer on Mal's Accounting OS platform. I own the AI service that ingests financial information on the platform from email triage and PDF extraction through GL classification, tax-code resolution, and double-entry journal generation. Python 3.12 · FastAPI · LangChain · Vertex AI (Gemini, `text-embedding-005`) · BigQuery vector search · Cloud SQL · Cloud Run · Terraform. - Built the financial documents extraction pipeline end-to-end: two-pass LangChain structured extraction against Gemini 2.0 Flash, a Promptfoo evaluation harness with per-field assertions, and OpenAPI-versioned output contracts. - Designed the confidence-based classification router that splits each line item into auto-accept / top-5 / LLM-fallback tiers using a similarity + agreement + gap formula, backed by a feedback loop (Postgres → Datastream CDC → BigQuery medallion → Vertex AI re-embedding). - Shipped the GDPR purge service: deletion registry, cascade across Postgres / BigQuery / vector store, Cloud Run + Cloud Scheduler deployment. - Built the feature-flag platform for trunk-based development — Pydantic flag schema, CI validation, zombie/expiry detection, multi-environment Terraform, gating every risky change across the codebase. - Own the full infrastructure surface via Terraform (10 modules across Cloud Run, Cloud SQL, BigQuery, Datastream, Pub/Sub, Vertex AI, VPC/PSC, IAP) deployed to dev / staging / prod.

- Built and delivered Media Planning Automation combining Generative AI with CatBoost ensembles (with HPO and stacking) to improve classification accuracy and decision support. - Implemented LLM observability and automated evaluation (Langfuse, Promptfoo), reducing evaluation cycles by 25% (saving ~6 engineer-hours per week). - Optimized unstructured data extraction using Gemini-family LLMs with chain-of-verification and structured outputs; cut ingestion latency by 52% (25s -> 12s) and increased key-field extraction accuracy by up to 60%. - Built ingestion and transformation pipelines across MongoDB, BigQuery, Google Sheets, and PDFs on GCP (Vertex AI, Cloud Run). - Co-developed a conversational research interface with GraphRAG and text-to-GQL for competitive intelligence; scaled ingestion of 145M JSON files into Cloud Spanner (graph) using Python concurrency for a 4.8x speedup (48h -> 10h). - Achieved 95% accuracy on text-to-GQL tasks using structured outputs, tool calling, and MCP. - Fine-tuned RoBERTa on a fake-news corpus for fact checking; reached 91% accuracy, 0.90 F1 (3-class), and 40 claims/sec on a single GPU. - Partnered with product and data engineering; mentored juniors, conducted interviews, and supported onboarding.

- Led an end-to-end credit card fraud detection pipeline augmenting existing antifraud systems; reduced manual investigation workload by 25%. - Evaluated GNN architectures (DGL, GTAN) against XGBoost baselines for transaction-graph fraud; shipped the winning ensemble into the risk pipeline. - Built ETL and feature engineering (PostgreSQL, Pandas, Feast) and automated training/testing/evaluation (XGBoost, Random Forest, DGL, GTAN). Shipped inference with Docker, Kubernetes, Metaflow, FastAPI. - Coordinated cross-functional engineering teams and external vendors using Scrum; delivered releases on schedule. - Integrated Cybersource Token Management Service; increased authorization rate by 3% and enriched behavioral data for downstream ML. - Integrated Cybersource Decision Manager into the risk pipeline (400k transactions/week); improved fraud detection effectiveness by 4.2%.

Joined Finstrument from an international payments BD role — my first full-time data science position, focused on fraud prevention and approval-rate analytics. - Co-led collaboration with Visa to stand up BI and a data-driven antifraud approach; increased approval rate by 6% in 6 months. - Built real-time observability dashboards (slicing/segmenting over 50 parameters, up to 5M records) with a Python/Pandas backend on Parquet; delivered Streamlit + Vega-Altair front end.

Novafile is a secure cloud storage service designed for modern world processes and work styles.