Oluwatosin – PyTorch, AWS, Python

A founder‑first business clarity platform that turns scattered metrics into simple, actionable guidance for first‑time founders. It emphasizes guided setup, plain language, and visual insights over complex dashboards, with multi‑tenant architecture, outcomes/goals tracking, scorecards, and an AI assistant (“The Operator”) for metric questions.

• Built a tool‑based chat pipeline that registers Operator tools, streams responses, tracks tool calls/results, and enforces required tool usage for data‑backed answers;
• Added system‑prompt loading from DB with a resilient fallback, plus automatic knowledge‑base context injection for concept questions;
• Centralized tool registration and instantiation for account‑scoped tool execution, supporting both RubyLLM native tools and legacy OpenAI function format;
• Implemented a “get_metrics_summary” tool that performs metric search, fetches recent time‑series summaries, computes trends, detects bottlenecks, and auto‑pulls knowledge‑base guidance for recommendations;
• Built a daily update generator that gathers revenue/vitals/goal status aligned with dashboard logic, assembles structured context, and produces a formatted update via LLM prompts;
• Added tool‑call monitoring, reasoning step tracking, UI broadcasts, and a “direct response” bypass for daily updates to guarantee formatting and avoid double responses.

Production “medical AI infrastructure” stack: a GPU-accelerated, evidence-grounded medical intelligence backend + web app that serves a medical agent. It runs as a FastAPI API with a React/Vite frontend and externalizes state to Redis + Postgres, connects to vLLM for LLM inference, and uses retrieval over FAISS plus Qdrant.

• Built a hybrid RAG pipeline over FAISS (PubMed abstracts) + Qdrant (medical codes) with parallel retrieval + prompt assembly, feeding vLLM for grounded generation;
• Implemented CIVERA repository-first guardrails that classify evidence completeness and dynamically constrain LLM outputs to retrieved sources (strict mode to avoid unconstrained generation);
• Workspaces monorepo migration (refactor/monorepo). Reorganized into src/api (FastAPI), src/web (Vite/React), infrastructure/docker, scripts/, tests/, and organized Docs/;
• DevX and ops hardening via unified Docker Compose workflow + health checks; standard dev scripts; structured test layout (unit/integration/e2e markers); clearer “how to run” docs and architecture handoff.

Developed a Generative AI–powered bookkeeping agent that automates the extraction of accounting entries, assigns GL and tax codes, and performs bank reconciliation. The system processes around 10 invoices every 3 minutes and supports multiple communication channels, including WhatsApp, email, and Slack. Designed for accounting firms, SMEs, and startups, the solution delivers high efficiency—up to 10× faster and 4× more cost-effective than traditional bookkeeping workflows.

• Led the development of an AI-powered accounting agentic system and designed end-to-end AI/ML architecture;
• Built scalable document-processing infrastructure handling tens of thousands of invoices monthly, including a custom VLM that outperformed traditional OCR;
• Developed a multi-agent system for accounting tasks using open-source foundation models and implemented RAG/GraphRAG pipelines for accurate extraction and reasoning;
• Optimized and scaled LLM/VLM inference using Kubernetes, GPU multi-LoRA adapters, quantization, flash attention, and CUDA-optimized engines (vLLM, LMDeploy);
• Designed ETL and automation workflows with Airflow for continual fine-tuning and model improvement;
• Created a custom orchestration layer (replacing LangChain and others) and built an in-house vector database for semantic search;
• Delivered a production transactional accounting agent using React, Python, FastAPI, and Slack SDK, deployed via AWS ECS, Docker, and Terraform;
• Established MLOps practices, CI/CD pipelines, observability (Grafana, Elasticsearch), and grew the engineering and AI teams from the ground up;
• Architected and deployed Label Studio and Argilla for structured annotation workflows automation across the AI team;
• Integrated Labelbox for enterprise-grade labeling at scale;
• Enabled a human-in-the-loop feedback loop that fed directly into Airflow-orchestrated fine-tuning pipelines;
• Managed data annotator sourcing and management;
• Reduced dependency on external annotation vendors by bringing labeling fully in-house over time.

RLHF preference pair annotation for code and mathematics domains, contributing to foundation model alignment and fine-tuning pipelines for one of the world's leading AI data companies.

• Evaluated and ranked model-generated code and math outputs to produce high-quality preference pairs for LLM fine-tuning;
• Contributed to alignment pipelines used in training state-of-the-art foundation models;
• Developed domain judgment for code correctness and mathematical reasoning at scale.

AI-powered contact center platform that analyzes conversations across digital and voice channels, delivers real-time coaching to advisors, and provides leaders with actionable insights into 100% of customer demand and performance metrics.

• Built and deployed transcription (Speech2Text wav2vec-large-xlsr model) technology for optimizing call center productivity for client queries;
• Used SQL best practices to design PostgreSQL back-end databases for insights;
• Streamlined ETL processes, reducing errors due to manual data entry;
• Built a multi-tenancy speech text classifier model using the Transformer DistilRoBerta model for client call center data, which were poorly mixed up during database setup for multiple customers' environments in Client Server;
• Built and deployed a continuous customer prioritization algorithm for managing customer churn rates through continuous training and inference using Ludwig and Ray clusters for a state-of-the-art MLOps Pipeline;
• Performed QA analysis of data-centric frontend product features, such as metrics, filters, and weights of call center agent performance with customers by querying MongoDB and transformation using pandas and numpy.

AI-driven platform for assessing news article credibility and fact-checking. The system uses transformer models to detect article stance and credibility, incorporates an evidence retrieval pipeline for verification, and includes automated CI/CD workflows for continuous model updates and deployment.

• Developed and deployed a news article credibility algorithm using DistilRoBERTa, applied commercially for fact-checking major events such as the 2020 US presidential debate and monitoring misinformation during Euro 2020;
• Built a stance detection model to identify agreement and disagreement, mitigating coordinated attacks and manipulative threats;
• Created an evidence retrieval system combining BM25 ranking and semantic search with embeddings, powering the fact-checking pipeline by sourcing claims from media and social platforms;
• Implemented CI/CD pipelines for continuous machine learning deployment on GCP;
• Developed an aspect-based topic modeling system to extract and generate topic terms from social media content, enhancing data analysis and monitoring.

AI-powered assistive technology solutions designed to support inclusive education and enhance independence for individuals with visual impairments or reading disabilities, such as dyslexia and ADHD. The core product includes a mobile app and smart reading glasses to facilitate accessible reading and learning experiences.

• Built a rule-based Clinical Decision Support system to detect early signs of blindness;
• Developed multisensory deep learning algorithms for embedded smart reading glasses;
• Implemented Optical Character Recognition (OCR) for real-time text capture;
• Created Text-to-Speech models using RNNs to provide audio output for visually impaired users;
• Integrated hardware and software components to optimize the performance of smart reading devices;
• Conducted testing and fine-tuning to ensure accessibility, accuracy, and responsiveness of the system.