Soma · Software engineer · United Kingdom
AI engineering·Agentic systems·Software engineering
I'm Soma, a First Class Computer Science graduate. Give me a slow, manual or messy process and I'll find the knot, build the fix and prove it works. I've done it across consulting, startups, academia and enterprises.
How I work
The same approach whether it's a charity's stock room, a hiring process or a national road network.
I start with the people doing the work, not the technology. Where does their time go? What keeps breaking?
At Projekt Rising that meant mapping how three businesses actually ran before writing a line of code.I pick the change worth the most and prototype it fast, so everyone is reacting to something real.
At SalesAPE I took 3 prototypes to leadership. 2 were greenlit.Then I build it properly, end to end: the data, the backend, and the part people actually click.
For Samaritan's Purse that meant replacing a spreadsheet with an order form, payments and a live stock dashboard.I don't call it done until the numbers say so: tests, evaluations and real runs.
The Ventureship hiring pipeline went through 3 rounds of production test runs before it was trusted.Projects
Each one starts as a problem, in grey. When it reaches you, it turns into what I did and what changed.
The knotAn AI hiring platform still ran hiring as a fully manual process, done by the founder.
What I didBuilt an end-to-end pipeline with the Anthropic SDK: multi-stage agents with self-checking schemas, deterministic decision logic, and bias audits at key decision points.
What changedOne automated pipeline run replaces the manual process, hardened over 3 rounds of production test runs.
The knotEngineering knowledge was spread across thousands of technical PDFs that constantly reference each other.
What I didExtracted and cleaned the text, built a pipeline to spot document codes and references, and turned it into a knowledge graph with a RAG assistant on top.
What changedStructured answers from thousands of linked documents, supporting infrastructure asset management.
The knotMost projects are a config file pointed at someone else's library. I wanted to prove I understood the machinery underneath.
What I didWrote a transformer from raw CUDA kernels up, with my own autograd engine and no deep learning framework in the core maths, then trained the same model on text and on movie histories.
What changed25.7x faster attention, and it runs live as a movie recommender.
The knot£10,000 of discounted shoebox gifts had to reach 50 community groups, and every order ran through a spreadsheet.
What I didDesigned and built the whole system: an online order form, automatic payment and confirmation, and a live stock dashboard that flags low stock.
What changed£10k of products and 50 groups handled in one place, built to cut admin time and order errors.
The knotCrash investigators need to know exactly which car was involved, and how safe it is, often from a single photo.
What I didBuilt a two-branch ResNet50 model that predicts make, model and year with calibrated confidence, matches it to Euro NCAP safety ratings, and returns a court-admissible report.
What changedAn answer in 46ms per image at 81.71% hierarchical accuracy.
The knotDeclined card payments get retried with the same rules for every bank, so retries are wasted and good sales are lost.
What I didModelled the full merchant to issuer flow, including routing, 3DS checks, retries and late chargebacks, then learned how each issuer behaves and tuned retries and routing to match.
What changedApprovals up from 82.5% to 84.3%, wasted retries cut from 8.1% to 4.2%, about £26,500 net benefit in simulation.
The knotEvery new CRM or accounting tool a client used meant writing another integration by hand.
What I didBuilt an engine that reads a product's API spec, ranks the useful endpoints, and generates a working client at runtime, all driven from a CLI.
What changedA new CRM now takes one adapter file. 146 passing tests, working with HubSpot and Pipedrive.
The knotRain, fog and snow make roadside CCTV hard to read, which weakens road monitoring.
What I didTrained CNN models on 10,000 CCTV images to recognise the weather in each frame, and a model to remove it from the picture.
What changed98% classification accuracy, making road monitoring more reliable.
The knotChecking that a face matches an ID photo is slow and error-prone when a person does it by hand.
What I didFounded an identity verification product and built it on a Siamese neural network, with TensorFlow and OpenCV pipelines tuned for accuracy and speed on real-world photos.
What changed2 paying clients and £5k in revenue.
The knotSmall service businesses need a website, lead capture and bookings, but rarely have the time to set them up.
What I didBuilt a full-stack platform in TypeScript, React and Postgres, with background job queues, that turns an Instagram profile or website link into a working online presence.
What changedA live site with lead capture and bookings, in minutes.
The knotStudents' notes, slides and recordings sit in files they can't easily question or revise from.
What I didBuilt a Django and React platform where you upload your materials, chat with them, and get quizzes and flashcards, with the heavy processing running in the background on Celery and Redis.
What changedSearch by meaning, chat and revision from your own materials, in one place.
The knotTwo very different briefs: make survey data collection feel less like a chore, and train an agent to win at Street Fighter III.
What I didBuilt a voice agent that turns a survey into a real-time conversation, and a reinforcement learning agent trained across 3 difficulty levels over about 10 million steps.
What changedTop 10 of 200 teams across Europe, and top 5 of about 50 teams.
Experience and awards
Nine roles, five awards and a First Class degree, from a national road network to a founder's first hire.
Founding Engineer Intern
The knotHiring on an AI hiring platform ran as a fully manual process the founder did by hand.
What I didBuilt an end-to-end agentic pipeline with the Anthropic SDK: multi-stage agents, self-checking schemas, deterministic decisions and bias audits at key points. Helped shape the shared data schema, reviewed weekly with the technical advisor.
What changedOne automated pipeline run replaces the manual process, hardened over 3 rounds of production tests.
Software Engineer, volunteering
The knotA £10,000 gift procurement project for 50 community groups was run from spreadsheets.
What I didDesigned and built the ordering, payment and stock system end to end, from the order form to the live dashboard.
What changed£10k of products, 50 groups, one system.
Head of Engineering
The knotA nonprofit focused on responsible AI governance needed someone to lead its engineering.
What I didLed engineering and managed a cross-functional team.
What changed12 months leading the team.
AI Engineer, contractor
The knotThree small businesses were losing hours every month to manual work.
What I didMapped their workflows to find the automation gaps, then built and monitored production APIs connecting models to their data, tracking accuracy, latency and outcomes.
What changed3 processes automated, about 60 hours of manual work saved every month.
Software Engineer Intern
The knotLeadership needed to know which new product ideas were worth backing.
What I didDesigned and built 3 MVP prototypes in a fast-moving hacksquad, with evaluation pipelines to measure how reliable they were.
What changed2 of 3 greenlit for further development.
Data Science, Year in Industry
The knotCCTV that failed in bad weather, knowledge buried in thousands of documents, and roadside habitats that needed mapping.
What I didBuilt a weather classification and removal model, a RAG knowledge graph over engineering documents, and a satellite segmentation model that classifies roadside habitats.
What changed98% accuracy on 10,000 CCTV images, and Year in Industry Student of the Year, 1 of 10 from 400.
Founder
The knotIdentity checks were slow and manual.
What I didFounded and built an image verification SaaS, from model training and inference pipelines to the product clients used.
What changed2 paying clients, £5k revenue.
Research Intern · Software Engineering Intern
The knotClimate resilience research, and credit risk analysis under strict financial data rules.
What I didAt Imperial, prototyped data-driven climate resilience technology with researchers and industry mentors. At Paradigm, built credit risk tools, client portfolio reports and internal web apps.
What changedNamed Paradigm Finance's Intern of the Year.
BSc Computer Science · First Class Honours
What I studiedAI, Data Science, Advanced Algorithms, Software Engineering and Theory of Computation.
Before thatCanadian Plus Foundation at 16 Plus Schools: 91% overall, and Best Foundation Student, top of 50.