● Contributing as one of three core developers building and scaling Tyce’s AI-powered document assistant. ● Implemented the ”Tyce Agent” using PydanticAI, including fine-tuning agent behavior, designing the agent tools, streaming responses, and optimizing. ● Optimized agentic RAG search by implementing contextual retrieval. ● Implemented custom HTML splitter for more stable and style-aware document chunking. ● Scaled data ingestion pipeline to handle 100,000+ in a few minutes by moving ingestion to AWS Lambda. ● Used advanced techniques, including chunking optimization, parallel processing, and selective processing, to cutdown processing costs and time by 20% - 90%, depending on the document. ● Participating in product roadmap and architecture discussions, helping shape both technical direction and user experience. Projects: 1. Movie Recommendation System — Python, LLM (Mistral AI), FastAPI. ● Designed & built a chat-based LLM movie recommender using Python, Hugging Face Transformers, Mistral-7B, LoRA/PEFT, and SQLite (FTS5 + vector blobs); ingested >1 M TMDb + IMDb titles with plots, genres, and embeddings for low-latency semantic search and sophisticated multi-filtered recommendations. ● Exposed the catalogue through an Mistral-based function-calling API with FastAPI, letting the model auto-fill JSON queries and rank candidates. ● Containerised with Docker and deployed inference on GroqCloud, streaming >3 k tokens/s; added token-level logging, cost monitoring, and CI/CD scripts for automated model refreshes. 2. Veritas - Android E-Book Reader — Kotlin, Firebase, GCP. ● Built MVP in 90 days; ran UX study with 20 pilot users. ● Added AI-based book recommendations increasing reading session length 20%. 3. Phonon - Cross-Platform Image Viewer — Electron, Node.js. ● Developed high-performance viewer (<200 ms cold start, <50 MB RAM) for Win/Linux/macOS. ● Adopted by friends for design workflows; praised for speed and minimal UI. 4. AI Image Colorization Tool — TensorFlow, OpenCV. ● Co-developed model colorizing 100 high-res landscape images per minute with 30% higher SSIM than baseline.

● Architected data pipelines that curated and transformed 10 000+ multimodal samples (2M+ tokens) for training state-of-the-art LLMs using the latest training formats (RLHF, SFT, DPO, ORPO). ● Fine-tuned and evaluated code-generation and chat models with Hugging Face Transformers, raising pass-@-1 accuracy to 70% and cutting hallucinations by 18%. ● Built an automated evaluation harness that reduced model validation cycles by 40%. ● Rolled out semi-supervised active-learning workflows and auto-labeling pipelines that slashed manual QA effort by 5x. ● Led a cross-functional team of 12 data engineers, mentoring them on prompt engineering, dataset design, and scalable MLOps best practices. ● Implemented robust PII and toxicity detection in Python & SQL, achieving 99,2% precision and ensuring SOC 2 / GDPR compliance. ● Deployed GenAI services on Kubernetes with GitOps practices using GitHub CI/CD pipelines, streamlining production rollout and rollback. ● Integrated models with external platforms like ServiceNow to enable real-time AI-assisted workflows. ● Provisioned cloud infrastructure on AWS, GCP, and Azure using Infrastructure as Code (IaC) to deploy and manage GenAI training utilities-including evaluation harnesses, data pipelines, and model serving-while maintaining a 99,99% SLA.

● Played a key role in developing and maintaining robust cloud-based ETL pipelines for a leading fashion analytics platform, ensuring a reliable flow of 4 TB+ raw data daily. ● Collaborated with a fully remote team across 3 time zones, leveraging sprint rituals and clear documentation to hit 95% of roadmap milestones on schedule. ● Re-engineered the extraction component of a fashion analytics pipeline (20k+ lines), shrinking daily runtime by 36% and saving $45k per year in compute. ● Built and maintained full-stack features: React dashboards for merchandising teams, backed by Django REST API endpoints and AWS API Gateway. ● Ensured data integrity and performance with automated tests (PyTest + Cypress), reaching 85% coverage and daily CI runs. ● Deployed micro-frontends and Python services via Docker/Kubernetes and Terraform, enabling blue-green releases with <5 min rollback.

● Delivered 100+ contracts ranging from predictive analytics to indie games. ● Built real-time analytics platform on GCP handling 60k events/sec; reduced query latency from 2 s to 200 ms. ● Optimized PySpark pipelines, lowering cluster costs 22% through improved partitioning.

● Improved API response time 55% and boosted uptime to 99,8% by refactoring Django ORM queries. ● Migrated the website’s hosting and database from shared cPanel to AWS EC2 and Amazon RDS, improving reliability and scalability.