A few that taught me something worth keeping.
Problem. ML models make claims on private data with no way to verify the inference without exposing the inputs or weights.
What I took away. Built a zero-knowledge ML trust layer using EZKL to compile an autoencoder to a ZK circuit. Deployed a Solidity reputation oracle on Ethereum Sepolia that publishes tamper-proof risk verdicts on-chain.
Problem. AI agents when asked for educational reflections, don't capture what actually happened in a session — the real mistakes, the wrong turns, and the corrections that actually taught you something.
What I took away. Built an end-to-end event-driven developer tool that turns noisy AI debugging traces into structured learning artifacts using lightweight session hooks, graph-based compression, and templated rendering. The project strengthened my ability to design for latency, context limits, and reliability by decoupling event capture from heavier reasoning, while reinforcing best practices in schema design, CLI architecture, local-first plugins, testing, optional NLP enrichment, and cross-platform AI tool integration.
Problem. Clients needed a smarter way to find and reach out to potential business leads — manually searching directories and crafting cold messages was too slow and inconsistent.
What I took away. How to build a full pipeline from Places API discovery through LLM-generated personalised outreach, and how much prompt tuning matters when the output goes straight to a client inbox.
Problem. Insurance fraud often hides inside large volumes of accident and claims data that are fragmented, inconsistent, and difficult to investigate manually. Without a reliable way to detect suspicious patterns early, high-risk cases can be missed while investigators spend time reviewing low-value ones.
What I took away. Learned how to transform messy, heterogeneous data into an explainable fraud detection workflow by combining domain rules, anomaly detection, and relationship analysis.
Problem. Manual processing of financial records and invoices is slow, error-prone, and inefficient, making it difficult to accurately extract and manage key financial data.
What I took away. Ongoing
Problem. Assessing the physical condition of used books manually is subjective, time-consuming, and inconsistent, especially at scale for resale platforms or libraries.
What I took away. Applied computer vision to real-world data, handled noisy and imbalanced datasets, optimized model performance through hyperparameter tuning and deployment formats (ONNX, TensorRT), and gained experience in data preprocessing and evaluation.
Problem. An investment committee evaluating a Tunisian AgriTech startup's Series A needs to query a dense, heterogeneous document corpus — financials, legal statutes, market studies — accurately and without hallucination, citing exact figures and refusing when information is absent.
What I took away. Retrieval quality matters more than model size: hybrid search (semantic + BM25 + RRF), semantic prefixes, and strict confidence thresholds keep outputs grounded. I built an event-driven tool that converts noisy AI debugging traces into structured artifacts via lightweight hooks, graph-based compression, and templated rendering. The Python/LangChain version offers robust persistence and hybrid retrieval, while n8n favors simpler deployment over depth.
Problem. Drafting emails and meeting summaries from rough bullet points, turning them into something readable without losing the writer's tone.
What I took away. Fine-tuning an LLM on business communication data, and the gap between raw model output and something a real person would send.
Problem. I built IRMusicSync to close the gap between passive LED strip remotes and genuinely audio-reactive lighting by turning an Android phone into a real-time bridge between live microphone input and infrared color control.
What I took away. I learned how to design a low-latency signal-processing loop on mobile by sampling PCM audio, extracting spectral energy across bass mid and high bands, detecting beat intensity from FFT-derived features, and mapping that analysis to stable IR commands that feel musical instead of noisy.
Problem. Predicting install counts across a 3M-row dataset of Play Store apps, and prototyping a market-research tool for indie developers.
What I took away. Handling scale with pandas and Keras without the notebook falling over, and how much cleaning matters before the model sees anything.