AI & safety
Model behaviour, evaluation, interpretability, and classical machine learning.
- PyTorch
- TensorFlow
- Transformers
- scikit-learn
- PyRIT
- RAMPART
Yuvan Aarya Chikka
Computer Science student at Queen Mary with over two years of hands-on AI engineering experience. Building AI systems, Python backends, and fullstack products from research prototypes to deployment.
Capabilities
Technical profile
Python-first engineering across AI experimentation, backend systems, and the interfaces that make them useful.
Multimodal systems
Retrieval and generation across video, images, speech, and text.
- LangChain
- CLIP
- Whisper
- spaCy
- Ollama
- FFmpeg
Python backend
Production APIs, relational services, and business-process automation.
- Django
- FastAPI
- Flask
- SQLAlchemy
- Twilio
- Redis
Data & modelling
Data pipelines, forecasting, analysis, and visualisation.
- PostgreSQL
- MySQL
- SQLite
- BigQuery
- Pandas
- NumPy
- Plotly
Fullstack delivery
Accessible web products, dashboards, and interactive interfaces.
- JavaScript
- HTML / CSS
- Google Maps API
- Folium
- Java
- C / C++
Cloud & DevOps
Cloud deployment, containers, CI/CD, and production operations.
- AWS
- Azure
- Google Cloud
- Docker
- GitHub Actions
- Git
- Linux
In practice
Experience
Jan 2025 - Present
Founding AI Engineer
1soul.ai
Developing the core WhatsApp experience through a FastAPI backend and Twilio, including bot conversation logic and delivery. Built persona representations, update rules, and custom environments for simulated persona-to-persona interactions, with CI/CD through GitHub Actions.
- Python
- FastAPI
- Twilio
- AI agents
- GitHub Actions
Jun - Aug 2025
AI + Data Analytics Intern
SAV Group
Built Django lead-management dashboards that unified messaging services and Monday.com, plus area-based pricing scripts used to evaluate planned acquisitions. Delivered an AWS-hosted RAG assistant for fiftyonelondon's web and WhatsApp sales workflows, led the development stream, and documented each tool with usable GUIs.
- Python
- Django
- SQL
- AWS
- Monday.com
May 2024 - Aug 2025
Founding AI Engineer
ReadyTraderOne
Built multi-asset order-book simulations covering crypto, FX, fixed income, and commodities, alongside a Django and Redis interface hosted on EC2. Designed educational market scenarios and trading agents using grid and mean-reversion strategies, qualitative news, and chat-driven market influence.
- Python
- Django
- Redis
- AWS
- Trading systems
Sep 2024 - Mar 2025
AI Engineer
Propmoolah
Reduced WhatsApp interaction time by over 60% by replacing menus with a FastAPI and Twilio NLP-to-SQL flow. Automated government, Rightmove, and Zoopla data into BigQuery, created time-series price forecasts, and built a sourced urban-planning document workflow that helped unlock investment and the company's relocation.
- Python
- FastAPI
- Twilio
- BigQuery
- Time series
Jan - May 2024
AI Engineer
Venu.AI
Raised scraping success by 15% with custom lead-aggregation automation, then improved outreach quality through automatically selected and filled email templates. Designed an Azure pipeline to scale processing across a cluster with stronger data validation.
- Python
- Web scraping
- LLMs
- Azure
- Automation
Aug 2018 - Feb 2023
Fullstack Engineer
Freelance
Built fullstack websites for a range of freelance projects, primarily Django-based internal inventory-management systems. Worked end-to-end across application logic, database-backed inventory flows, and browser interfaces.
- Python
- Django
- JavaScript
- HTML / CSS
- SQL
Foundation
Education & certifications
Queen Mary University of London
BSc Computer Science
Sep 2025 - Jun 2028. Previously earned four A*s at A-level in Mathematics, Further Mathematics, Physics, and Electronics.
84
% current GPA
Professional development
Professional certifications
Industry-recognised certification and focused programmes spanning applied AI, technical AI safety, deep learning, NLP, mathematics, and data science.
- Amazon Web Services Industry certification AWS Certified AI Practitioner
- BlueDot Impact Technical AI Safety
- Coursera Deep Learning Specialization
- Coursera Natural Language Processing Specialization
- Coursera Mathematics for Machine Learning and Data Science
- IBM Data Science Professional Certificate
- Harvard University CS50W: Web Programming with Python and JavaScript
Research & selected builds
Projects
Research into model behaviour and practical systems for multimodal retrieval and accessible transport, presented with the work itself.
Mechanistic interpretability / Apart Research
Private Is Not Privileged
Tests whether activation-based predictors have a genuine information advantage over matched external baselines. In a 40-game Diplomacy experiment, an apparent white-box advantage reversed once the text observer received matched post-deliberation information, with robustness audits across model families, timings, and reader classes.
- 40Diplomacy games
- 0.103activation Brier score
- 0.033matched-text Brier score
Mechanistic interpretability / Apart Research
GraftLoyal
Mechanistically constructs context-sensitive secret loyalty in a frozen language model by reading internal representations and intervening through a gated writer vector. Evaluated on 1,152 prompts across counterfactual pairs and held-out domains.
- 1,152evaluation prompts
- 0.991active-direction AUROC
- +1.39–3.89pp95% CI / aligned-choice lift
AI evaluation / Apart Research
RiskLab
A multi-evaluator framework for detecting manipulative behaviour in language models. It combines rule-based analysis, classifiers, and councils of judges while preserving provenance instead of trusting a single evaluator.
- 150+scenario library
- 20evaluated episodes
- 0.014maximum mean framing delta
LongVideo RAG
Fully local question answering over long videos using transcription, visual frame retrieval, timestamped evidence clips, SQLite indexing, and Ollama generation.
TransitAssist
A Django public-transport planner with interactive routing and community reporting for accessibility barriers and successful accessible journeys.
Selected signals
Recognition
Awards, leadership, and competitive results outside day-to-day engineering work.
Imperial AI Ventures
Built a video-interview system that generated questions from student essays and used gaze and sentiment analysis with a four-person team.
QMUL AI Society President
Elected to lead the university's AI Society and its programme of technical events, workshops, and student-led projects.
Voice AI Hack
Co-organised a one-day London hackathon for up to 60 builders across Voice & Medical and Voice & Productivity tracks, bringing together industry judges, sponsors, and technical teams.
Event page ↗Technical competitions
TMUA 8.0, USACO Silver, top 10% in the Europe Reply Challenge, and qualification across British mathematics competitions and olympiads.
Contact
Find me elsewhere
Profiles, direct contact, and the full CV.