LLMs and NLP#
Language models and NLP put to work where the output has to be measured: recommenders, classifiers, speech evaluation, retrieval and verifiable inference. Six projects, 2021 to 2025.
AI in smart contracts for decentralized finance#
Problem · Smart contracts can act on the outputs of AI models. How do you verify that an inference API's output is trustworthy and came from the model you are paying for, rather than a cheaper one or a fabrication, without re-running it?
What I did · Contributed to a verification framework that authenticates machine learning predictions with cryptographic fingerprints at key inference stages, demonstrated the prover and verifier protocol on ensembles, agentic workflows, streaming and batched inference, and built verifiable price-prediction endpoints. Later work added LLM-driven trading-strategy selection.
Result · The protocol was published as a preprint and released as an open-source reference implementation; see Research.
Stack · Python, cryptographic fingerprinting, LangChain, prediction APIs.
CEO
"Would you like to join us as Chief AI Scientist?"
Voice-to-text agent benchmarking#
Problem · Voice agents depend on accurate speech-to-text across languages, and the hard cases are alphanumerics: booking codes, phone numbers, spelled-out digits. Comparing quality across locales is hard because numbers, dates and names are written differently, and some scripts need romanization first.
What I did · Built a multilingual evaluation framework for a conversational AI company: import of public speech corpora, cloud transcription, word and character error rates with median and tail statistics, inverse text normalization, romanization, and an entity-level score for the alphanumerics that matter. Generated synthetic alphanumeric utterances with text-to-speech and used them to fine-tune custom speech models.
Result · The entity score exposed where grouped digits failed while separated digits passed, and drove the fixes. The framework covers many locales with comparable results across them.
Stack · Azure Speech, Python, SSML, jiwer, NeMo text normalization.
Reusable GenAI components for a global pharma company#
Problem · One-off GenAI prototypes were piling up in the operations IT of a global pharmaceutical company. It needed a clean library of reusable components instead.
What I did · Built a reusable retrieval-augmented generation backend on AWS Bedrock knowledge bases and agents, with a React frontend and a component library, developer and user guides, and a design for a reusable text-to-SQL component that favored plain REST over bespoke protocols so other teams could adopt it.
Result · Components and documentation handed over for reuse across teams.
Stack · FastAPI, AWS Bedrock, Docker, React, Azure sign-in.
GraphRAG for legacy codebases#
Problem · An AI developer-tools venture wanted to know whether AI can plan changes and refactorings in large legacy codebases before investing in a product.
What I did · Built retrieval over a code graph: a Java parser feeding an abstract-syntax-tree graph into Neo4j, a question benchmark on a reference codebase, code-quality metrics, and a refactoring planner.
Result · Answer accuracy was high on clean code and dropped sharply once the code was made cryptic, a finding that shaped the venture's product direction.
Stack · Neo4j, FastAPI, Streamlit, Java AST parsing, LLM APIs.
Recommendation systems for researchers#
Problem · Researchers must work out which journals and collections suit their work, out of thousands. An academic publisher is in a unique position to recommend.
What I did · Consolidated and cleansed researcher data, built a family of SQL-based probabilistic recommenders and the endpoints that serve them, a dashboard for the team, and upskilled the client's engineers to run and extend the system. The latest version combines submission history, fine-tuned LLMs, vector embeddings and statistical models.
Result · Recommendations serve millions of researchers across hundreds of community pages. Algorithm improvements raised click-through by about a third.
Stack · BigQuery, Firestore, Dataflow, Kotlin, Terraform, Looker Studio.
Head of Product, Researcher Profile
"Would you like to join us full-time?"
Text classification for regulatory compliance#
Problem · Local planning authorities in the UK receive thousands of written representations per consultation, over several rounds per plan. The volume overwhelms both the authorities and the national inspectorate that examines the plans.
What I did · Ran a discovery with inspectors and planning authorities, from mock-up voting to feasibility tests on documents from several councils, then built a segmentation and classification prototype with fine-tuned language models and entity and topic extraction, and a demo app for both sides. Policy names proved better categories than plan headings.
Result · In simulations, correct predictions rose from roughly six in ten to nine in ten as verified labels grew, saving weeks of staff time per consultation.
Stack · Transformer classifiers, spaCy, Python, Voilà.
Senior Data Scientist, UK government planning body
"We really got the maximum value for money out of this project."
Working on something similar?
The quickest way to find out whether I can help is a short call. Prefer email? Write to [email protected].