AI Engineer · World Trade Organization · Geneva

AI that holds up in production, and in an audit.

I embed with the teams who own a problem and take AI from first prototype to a system they rely on every day, built on their infrastructure with governance in place.

A query finding its closest matches: the core move behind every retrieval system I build

Selected work

Four AI systems now in production at the World Trade Organization

All work

16+

sources monitored continuously

Public sector · Trade policy

Monitoring the world's new trade measures continuously, not by hand

The World Trade Organization now tracks new tariffs, export restrictions and subsidies across 16+ sources automatically, with 17% lower AI model usage, and analysts review findings instead of hunting for them.

lower AI model usage, by removing duplicates first
17%

100%

of AI activity logged and auditable

Public sector · AI governance

Running AI on the WTO's own servers, with every output on record

The World Trade Organization handles sensitive information. It now runs its language models, search and storage entirely in-house, with every AI interaction logged for audit and one standard way to bring new models online.

external AI providers receiving the organization's data
0

97%

in testing: the right code in the top five

Public sector · Trade classification

Classifying products into customs codes in under three seconds

Member governments describe a product in plain words and get a shortlist of five HS codes in under three seconds. In testing, the right code was on that list 97% of the time, and usually at the top.

from product description to shortlist
<3 s

How I work

I embed with the people who own the problem, and stay until the system runs without me

  1. Embed

    Work alongside the people who do the job, and map the decision and workflow before writing any code.

  2. Prove it

    A working prototype on real data, so the case is proven, or the project stopped, before it gets expensive.

  3. Ship it

    A production system on the client's infrastructure, with evaluation, monitoring and governance built in.

  4. Hand it over

    Procedures, documentation and training, so the team runs and extends it without me.

Experience

Three years shipping AI to real users, from airport travelers to member governments

6 months

from intern to AI Engineer

Mar 2026 – present

World Trade Organization

Intern → AI Engineer, Geneva, CH

Built the trade-measure monitoring prototype as an intern, then was promoted to AI Engineer within six months and took it into production.

What I did
  • Built the trade-measure monitoring prototype (MCP, Google ADK, FastAPI): 7 sources across 6 jurisdictions, replacing a fully manual workflow.
  • Took it into production with department heads and trade analysts and scaled it to 16+ sources. Removing duplicates before AI processing cut model usage by 17%.
  • Built a classifier over 10,000+ regulatory codes (hybrid retrieval) with 73.6% top-1 accuracy at under $0.01 per query, routing low-confidence cases to human review.
  • Shipped an HS code recommender for member governments: in testing, the right code is in the top five 97% of the time, in under three seconds, and it is ready for the next HS revision.
  • Deployed a knowledge assistant over 5,400+ internal documents that enforces each user's access rights on every query.
  • Benchmarked open models (Gemma, Qwen) on the WTO's own GPUs to choose the production model.
  • Part of the team that built the WTO's sovereign AI platform (vLLM, Weaviate, MinIO), so no data leaves its own infrastructure.
  • Part of the team that put organization-wide AI governance in place (Langfuse, guardrails), with all AI activity logged and auditable.

Visit World Trade Organization

15%

revenue growth for a UK airport parking client, supported by my demand forecasts

Jan 2023 – Aug 2024

Perimattic Inc.

AI Engineer, Bengaluru

Forecast parking demand for a UK airport parking client, supporting the dynamic pricing behind its 15% revenue growth.

What I did
  • Deployed conversational AI agents (LangChain) serving 1M+ airport travelers, integrated with legacy back-end systems at 99.95% uptime.
  • Delivered enterprise knowledge agents over Google Docs and SharePoint, giving non-technical staff plain-language access to internal knowledge.
  • Built an LLM evaluation framework (DeepEval) that blocks a release when quality regresses.
  • Forecast parking demand for a UK airport parking client. The forecasts underpinned dynamic pricing that contributed to 15% revenue growth.

Visit Perimattic Inc.

<5 s

to read a number plate at an airport, down from about 60 seconds

Mar 2022 – Jul 2023

Freelance

Computer vision engineer, Jaipur

Built real-time number-plate recognition for an airport, from model to deployment.

What I did
  • Trained a YOLO detection model and deployed it for real-time recognition at the airport.

What I bring

Strategy, engineering and governance, so a system can be justified, built and trusted

01

Strategy

Find where AI pays back, and where it won't.

  • Use cases scoped and prioritized with department heads and the analysts who do the work
  • Stop-or-scale decisions on evidence, before a pilot becomes a sunk cost
  • Adoption: training, handoff and procedures teams actually use

02

Engineering

Build systems that run in production, on infrastructure the client controls.

  • Agentic and retrieval systems, from first prototype to daily use
  • Models benchmarked and served on the client's own hardware
  • Forecasting and pricing models tied to revenue

03

Governance

Make AI auditable and safe to scale.

  • Access rights enforced on every query
  • Release gates that block a release when quality drops
  • A full audit trail: every AI interaction logged

Toolkit

Agents and retrieval
RAGMCPLangGraphGoogle ADK
Models
PyTorchHugging FacevLLMProphet
Data and search
WeaviateQdrantBM25SQL ServerDatabricks
Delivery and cloud
DockerKubernetesCI/CDFastAPIAzureAWSGCP
Governance
EU AI ActGDPRLangfuseDeepEvalRAGASMLflowWeights & BiasesGrafana
Languages
PythonSQLTypeScriptC++

Let's talk.

I'm open to consulting and forward-deployed roles where AI strategy meets delivery. Email is the fastest way to reach me.