Leading the company-wide AI Lab: internal accelerators, rapid proofs of concept and standard tools.
Managing a seven-member team: delegation, mentoring and bench-resource activation.
Cut proof-of-concept delivery from 1 to 2 weeks to 1 to 2 days with agentic-coding workflows (Claude Code), a 5 to 10x speed-up.
More
Designing solution architectures and AI strategies for sales and executive teams, turning business needs into technical roadmaps and client pitches.
Working with senior leadership to set priorities, prototype solutions and show AI capabilities in stakeholder demos and technical presentations.
Driving AI standardisation through best practices, SOPs and reusable frameworks across AI and ML work.
Driving AI strategy for the SDLC: agentic-coding adoption across teams, an internal accelerator IP catalogue and AI-development process documentation.
Running knowledge-sharing and upskilling programs to build an internal pipeline of AI and ML expertise.
to Pune, India
Senior AI EngineerBizAmica
Built the first LLM products the company sold: four offerings across logistics and banking.
Cut deployment time from 5 days to 1 with Ansible, Docker and Jenkins.
Built distributed, GPU-accelerated AI and ML pipelines for 200,000+ documents a day; inference per document went from 6 to 7 seconds on CPU to about 1.5 seconds (about 78% less).
More
Became the primary point of contact for LLM expertise, working directly with VPs and CFOs on demos, pricing models and requirements.
Managed cloud and on-premises infrastructure with strict security controls in sensitive banking environments.
Trained 25+ employees and interns on API development, AWS, LLMs and DevOps.
to Leeds, UK
Data ScientistSky Betting and Gaming
Built an end-to-end ML solution for bonus-abuse fraud detection on about 100,000 monthly transactions, achieving 89 to 95% accuracy in production; deployed with Docker and Apache Airflow for the risk-management team.
Created custom prediction-threshold optimisation aligning model decisions with business priorities; proposed a novel data feature with comparative-modelling evidence projecting accuracy lift to 96 to 98%.
Client work is confidential. These projects are described, not demoed.
Projects from my roles and open source
Skill-routed ServiceNow agent
EU flight-catering client, IT service management
Replaced a failed deployment with skill prompts routed over an MCP tool layer. Latency went from 1 to 2 minutes to about 15 seconds.
More
Six modular skill prompts replace one mega-prompt with every tool attached. They cover incident creation, knowledge-base answers (RAG), service requests and ITSM FAQ, with dynamic category and location resolution. Working build in one week, production in two.
TokensPrompt tokens cut from 10,000 to 15,000 to about 5,000, about 66% less
CacheRedis semantic cache seeded with 50 to 60 common Q&A pairs; 60% hit rate, and common queries answer in 1 to 3 seconds
SecurityPII redaction, prompt-injection defence and content filtering for EU GDPR, on a fully private Azure network
ReliabilityAsync backend with rate limiting, circuit breakers and per-user token and cost tracking in App Insights
RolloutAzure DevOps CI/CD; piloting with 50 internal users, targeting several thousand
My role: architect, shipping with one DevOps engineer.
Azure
Azure AI Foundry
React
FastAPI
FastMCP
MCP
RAG
Redis
ServiceNow
App Insights
EDI 835 claim-payment automation
US healthcare client, Medicare
Automated EDI 835 payment processing on AWS Bedrock, at 95% UAT accuracy across the claim types the client enabled.
More
A feedback loop captures user corrections in production and feeds them back to limit drift. It now processes 20 to 30 batches a day. Cloud-native on Python, Docker, Lambda, S3 and Secrets Manager.
My role: owned discovery, architecture, delivery and documentation.
Python
AWS Bedrock
LangChain
Bedrock Flows
Lambda
S3
Secrets Manager
Docker
Lending and mortgage AI accelerator
US lending and mortgage client, Marlabs
A ten-track accelerator program: three tracks in production, seven on the roadmap, about two weeks to build each track.
More
Document extraction across all standard mortgage and lending documents on AWS Bedrock, with cross-document verification and external checks.
Lead scoring from website behaviour, with an LLM briefing for loan officers (hot or cold lead, talking points) and a drafted outreach email.
A lock-in extension agent that writes a recommendation brief from credit-score changes, new loans, profile changes and comparable past decisions.
My role: sole architect across all ten tracks.
AWS Bedrock
LLM
ML lead scoring
Document automation
SQL to Databricks migration agent
Data platform migration
A multi-agent migration: about 10 million rows in 20 minutes, against 6 hours by hand.
More
Pydantic AI agents handle schema analysis, conversion to Unity Catalog, documentation, S3-staged transfer and validation inside Databricks notebooks, with full lineage the manual process did not produce. The repeatable framework led to a follow-on Databricks to GCP migration sale.
My role: architect; grew the team from 2 to 7 developers.
Pydantic AI
Databricks
Unity Catalog
S3
Logistics and air-cargo AI suite
Freight forwarding and airport operations
Tonnage forecasting, an LLM-powered IVR and an airport operations chatbot.
More
Air-cargo tonnage forecasting by airport, cost centre, customer and commodity; the finance forecast cycle went from 6 weeks to minutes.
An LLM-powered IVR for routine customer inquiries; support calls down 30 to 40%.
A real-time airport and dock-management chatbot for freight forwarders and airport authorities.
LLM
Time-series forecasting
IVR
Chatbot
High-volume FASTag infrastructure
Indian banking client, vehicle KYV
AWS infrastructure for 14,000 FASTags a day plus a 7.5 million FASTag backlog, supporting the bank's Know Your Vehicle process.
More
Networking, load balancing, autoscaling, security and deployment.
My role: architected and deployed.
AWS
Load balancing
Autoscaling
Computer-vision fault detection
Automotive manufacturing, PLC assembly lines
Finds silent faults in PLC programming with computer vision and pattern matching: a 96% efficiency improvement with less assembly-line disruption.
More
Computer vision
Pattern matching
PLC
UCP 600 compliance agent
Open source, trade finance
Examines Letters of Credit presentations against UCP 600 articles and writes structured discrepancy reports, replacing manual bank examination.
More
Eleven parallel specialist sub-agents in a scalable, auditable workflow.
Classifies and extracts structured data from Bills of Lading, Commercial Invoices and Packing Lists, with confidence scores, cost tracking and audit logging.
More
The backend stays document-agnostic, for reuse across domains.