Consultant – Research Lead – NLP / GPT and Ontology Engineer

AI driven Whole Portfolio Optimization and buy side reseaech summarization.

Redesign, Rebuild, and Migrate the Semantic Engine supporting the Metadata for a Number of Datasources from the current third party tool to an in-house replacement

The project required to replace the current implementation of the semantic data hub. The solution required design a product able to handle a volume of metadata collected across across multiple divisions, develop a proof of concept (PoC), confirm the PoC with the stakeholders, and deliver a fully fledged implementation suitable for productionization. The solution consisted of a set of extractors, based on meltano, custom API connectors and ingestors written in Python to harvest the metadata from different sources. The metadata was then staged in Postgres and cleansed using DBT transformations. The Clean metadata got mapped to the internal ontologies using rmlmapper, transformed into triples and nquads, and loaded into Allegrograph.

Once in Allergrograph we used SPARQL queries to augment data across different graphs and extract Knowledge out of the bulk of information. The solution is designed to be deployed to AWS using a combination of native services (Airflow, S3, RDS-Postgres, EKS) and containers (Allegrograph, custom transformers, extractors, loaders, Meltano, DBT, RMLMapper). Final workload yielded datasets in excess of 50Mln triples.

Technologies: Stakeholder engagement, yarrml, RMLMapper, Allegrogarph, SQL, SPARQL, Graph Database, OWL, RDF, Ontologies, Protege, Meltano, DBT, Postgres, Snowflake,SnowPipe, Matillion, Python, FastApi, Flask, Django, GIT, GitHub, GitActions, CICD, CI/CD, Jira, TDD, DevOps, AWS, Terraform, Docker, Docker Compose, Airflow, RDS-Postgres, Cloud, AWS, API, sFTP.

Translate Data Science Models (R, Jupyter Notebook, Matlab) in production ready application in the Azure Cloud and on-prem Hadoop/Spark cluster

Project supported:

• Commodities trading project, It produced multi-million savings in the procurement of wheat and corn. Note we were trading on the bullish market leading to the Ukrainian crisis, thus the market situation might account for part of the performance.
• Product match: Automatically populating best match between internal and competitor product reducing manual intervention of a factor 4x, the average match time per item going down from about 3 minutes to less than a minute. The proposed item was accepted in 96% of the cases.
• Fresh: Modeling price reduction of products close to expiration date, preliminary results hint at a reduction in waste in the order of 20%.

Technologies: Stakeholder engagement, Java (EE), Python, Pandas, NLTK, Scipy, Numpy, Hadoop, HIVE, Pyspark, FastApi, Flask, Django, GIT, GitHub, Jenkins, Jira, CI/CD, TDD, DevOps, Automated Testing, Load Testing, ETL, Pipelines, Data Preprocessing, Data Lake, Azure, AzureML, Kafka, Spark, Hadoop, Hive, SQL, PostgreSQL, Teradata, Refinitiv Point Connect, Bloomberg SAPI.

Revamped the automated trade surveillance platform to meet the criteria set by the auditor (Team of 6 - co-located).

• Mediated between stakeholders to have them agree on a standardized approach across different asset classes.
• Mediated between stakeholders and developers to ensure delivery was meeting requirement.
• Defined templates for efficient and standardized implementation all analytic.
• Implemented a set of critical high-end analytic using NLP, ML, and advanced quant methods.

Initial challenges: Pending review from the regulator. The project suffered disconnect between stakeholders, compliance requirements, and developers. The platform was legacy. The development team suffered high attrition rate, thus loss of knowledge. Documentation was partial.

Benefits: Passing auditing (serious cost reduction). Providing meaningful alerts (67% spam reduction on downstream teams). Platform was consolidate and made extensible.

Asset Classes: FX spot/options, rates futures/bonds/swaps, repo, bespoke OTC.

Technologies: Stakeholder engagement, Java (EE), Python, Pandas, NLTK, Scipy, Numpy, Pyspark, Dask, Bitbucket, Jenkins, Jira, CI/CD, TDD, DevOps, Risk Scenarions, Automated Testing, Load Testing.

Delivered "Project James". A reinforcement learning AI for direct marketing optimization.

News UK won a Google sponsored innovation grant aimed at delivering an advanced solution to real marketing problems. Attrition of the initial investigator created condition for reassigning the task. The intervention required assessment of the partially implemented project, baseline the approach, rebuild the reinforcement learning core using state of the art tools. Tune and deliver a production viable tool within the scheduled time frame.

Challenges: Time pressure for delivery. Partially implemented platform with partial documentation. Full research project with no previous case study to leverage for comparison.

Benefits: "JAMES has revolutionised churn further, and advisors informed by readers interests underpin an award winning contact centre" ([link]

Technologies: Python, pandas, scipy, numpy, TensorFlow, Django, Flask, github, jenkins, jira, GitOps, CI/CD, DevOps, Kubernetes, Docker, Terraform, Microservice Architecture

Managed the delivery of the Cloud Logging and Monitoring Platform (Team of 20 across 3 sites).

In the framework of public cloud adoption, JPMC needed a standardized, large scale, logging and monitoring system to meet cyber-security requirements for all application in the public cloud.

Davide joined the team after the PoC of the platform. He reviewed architecture and implementation. Then, scaled the platform to handle 5TB of data a day (approximately 5 billion messages with peak of 1.3 bln during the first hour of trading).

Challenges: very new project, under hard constraints in terms of data protections, thus limited availability of approved cloud services. Very challenging requirements in terms of SLO/SLA, high availability, disaster recovery, and sustained recovery.

Benefits: The platform allowed to monitor an initial set of 5 mission critical applications in the public cloud (AWS). It pioneered new technologies, produced a number of architectural patterns new to JPMC, and demonstrated its ability to scale up to a higher number of monitored applications at a button push.

Technologies: Leadership, AWS (API Gateway, Route53, S3, DynamoDB, Kinesis, Elastic Beanstalk, Lambda, ELB, AIM, CloudWatch, CloudTrail, etc.), Boto, Terraform, FluentD, Kafka, Kafka Streams, (replaced by Kinesis after SOC3), Kinesis Firehose, NiFi, Elastic Search, Logstash, Kibana, Java (EE), Python, Bitbucket, Jenkins, Jira, CI/CD, TDD, BDD, DevOps, Hera (JPMC Terraform based API), Automate Testing, Load Testing. Microservice architecture, Docker, Kubernetes (k8s), Datadog. L1 and L3 support during rollout and production, respectively.

Developed the Meta Analytic of Corporate and Investment Branch (CIB) of the bank.

As part of the digital transformation initiative, JPMC aimed at labelling and scoring all data repositories and all software products owned by the line of business. We defined the Data Quality Metrics, formal ontologies for data representation of logical data models (LDM), scan through the meta-data of all databases inferring the physical data model (PDM) and linked them through heuristics. The results were manually refined by Information Architects.

Challenges: Very broad collections of dishomogeneous data. Data quality was not always prime. Some data steward was only partially cooperative with the process.

Benefits: The semi-automated approach increased productivity of the Information Architects by a factor 4.7x.

Technologies: Java, Spring, Python, RDF, OWL, Semantic web standards, Ontologies,Knowledge graphs, Graph Database, BigQuery, ISO11179, bitbucket, jenkins, jira, CI/CD, TDD, DevOps.