Building intelligent systems. Studying how they work. Sharing what I learn.

Retrieval-augmented generation is supposed to ground models in evidence. In medical vision-language models it can do the opposite. This talk walks through ongoing research on retrieval-induced hallucination in chest X-ray reporting: 95% of retrieval-augmented reports copied text verbatim from another patient's report (0% without retrieval), and the cause traced to embedding similarity tracking anatomy rather than disease. Retrieval doubled clinical accuracy when the retrieved case was relevant (CheXbert F1 0.201 → 0.402) and collapsed it to 0.043 when it wasn't. We cover the full pipeline — BioMedCLIP retrieval, LLaVA-1.5-7B generation, CheXbert evaluation — plus the patient-level leakage and coincidental-overlap controls that rule out chance, and what this means for anyone deploying RAG in a high-stakes domain.

Attendees leave with

• Why similarity search can silently retrieve the wrong evidence — and how to detect it
• A reproducible evaluation harness for grounding failures in medical VLMs
• Concrete gating strategies for when retrieval should be trusted, and when it shouldn't

AidCare is a clinical decision-support platform built for healthcare settings where specialist knowledge is scarce and infrastructure is unreliable. This is a systems talk about the engineering decisions that weren't obvious: why a dual-mode RAG design with separate knowledge bases for physicians and community health workers; how a Sentence Transformers + FAISS pipeline over 500+ clinical protocols delivers sub-second retrieval; how Whisper ASR and Tesseract OCR turn voice notes and scanned records into queryable context; and how a token-efficient Gemini prompting layer produces grounded differential diagnoses. We close with deployment on a single DigitalOcean box — containerized FastAPI, RBAC, scoped tokens, anonymized logging — and the failure modes we hit along the way.

Attendees leave with

• A reference architecture for domain RAG under real infrastructure constraints
• How to design ingestion for messy multimodal inputs (voice, scans, PDFs)
• Privacy and access-control patterns for clinical data on a small budget

Multi-agent systems are easy to demo and hard to make useful. This talk dissects a platform built for the Virtue Foundation that analyzes 797 healthcare facilities across Ghana's 16 regions: a LangGraph supervisor routes natural-language questions to six specialized sub-agents backed by a FAISS vector store, and an intelligent document-processing pipeline (GPT-4o-mini) extracts structured capability data from unstructured facility records with per-field confidence scores. It surfaced 10 medical deserts and 43 data anomalies — and took 2nd place at the Databricks × Hack-Nation Global AI Hackathon. We focus on routing design, when to split agents versus tools, confidence scoring for extraction, and how to keep an agentic pipeline debuggable.

Attendees leave with

• A decision framework for supervisor/sub-agent decomposition
• Per-field confidence scoring as a first-class output of LLM extraction
• Observability patterns that make agent failures diagnosable, not mysterious

Super-resolution looks better. Does it work better? Built in a three-person team at the ACVSS 2026 research hackathon, this project treats enhancement as an intermediate representation whose value must be proven on a downstream task: a three-phase pipeline (conditioned latent-diffusion SR → flood-aware fine-tuning → U-Net segmentation → per-tile risk ranking) on a SpaceNet-8 subset. Controlled experiments against degraded-LR and bicubic baselines showed diffusion SR improved every downstream metric (mean IoU +24%, flood mIoU +11%) — while an earlier SR run hurt segmentation, isolating semantic fidelity versus hallucinated structure as the condition under which enhancement helps. A talk about experimental discipline as much as about models.

Attendees leave with

• How to set up controlled comparisons that isolate what actually helps
• When generative enhancement hallucinates structure — and how to catch it
• A template for “does this preprocessing step earn its place?” experiments

Pediatric chest X-rays are acquired at low dose, so they're noisy — and denoising that blurs fine structure can hurt the downstream diagnosis it was meant to help. SharpXR, published at the MIRASOL Workshop at MICCAI 2025, benchmarks seven denoising baselines (REDCNN, DnCNN, HFormer, ResUNet++, Attention U-Net, Sharp U-Net, BM3D) and shows that structure-preserving denoising raises downstream pneumonia-classification accuracy from 88.8% to 92.5%. This talk covers the evaluation design — judging denoisers by task performance rather than pixel metrics — the architectural choices that preserve edges, and lessons from presenting the work at MIRG-ICAIR 2025 and the African Computer Vision Summer School.

Attendees leave with

• Why task-based evaluation beats PSNR/SSIM for medical image enhancement
• Which architectural choices actually preserve diagnostic structure
• A practical benchmarking protocol you can reuse on your own modality

A walkthrough of the paper that started the Transformer era, built for students and early researchers reading their first landmark paper. We break down self-attention, multi-head attention, and positional encoding with intuitive explanations and small worked examples — and, just as importantly, model how to read a dense ML paper: what to skim, what to slow down on, which assumptions to question, and how to check your understanding. Previously delivered to a student researchers session; slides available.

Attendees leave with

• An intuition for attention that survives contact with the equations
• A repeatable method for reading and dissecting ML papers
• Confidence to tackle the next paper alone

Twenty-plus hackathons with wins and podiums on three continents — HackZurich's Hybrid Team Award, GDG Lagos' ₦2M winner-takes-all Amala Hackathon, 1st runner-up among 800+ at HackLab Nigeria, 2nd at the Databricks × Hack-Nation Global AI Hackathon. This talk distills what actually transfers from a 48-hour build to a real product: scoping ruthlessly, choosing the boring stack, demoing the risky part first, and turning judges' questions into a roadmap. Practical, story-driven, and honest about the projects that didn't work.

Attendees leave with

• A scoping method that produces a working demo every time
• How to pick problems that judges — and users — care about
• The path from a weekend prototype to something people rely on