Ebenezer Tarubinga

Hybrid retrieval and reranking over a water utility's operations data, reasoning across a knowledge graph so every answer carries its evidence. Scored on a held-out set.

A tool-calling agent that reads the live grid strictly read-only, so it cannot act on infrastructure. Guardrails at every tool boundary, adversarial inputs in the eval harness.

Live motorway telemetry on a Three.js map over PostGIS, for an international smart-mobility programme. My part was hardening a prototype until it could pass certification.

On-device detection on municipal camera poles, MQTT and serial down to the hardware. Offline-first: it survives a cloud outage and recovers a device that stops reporting.

One OIDC provider answers two questions for every product: who is this request, and is this organisation's subscription active. Identity fails closed and the gate fails open, so a billing blip cannot lock out someone who paid.

Moved a fleet off managed Postgres onto one instance behind pgbouncer, one role per product. What forced it was an ORM distinct with no limit: 54,000 rows every five minutes to compute eight booleans.

Two free tiers chained, so no product holds a paid model key. It fails over on 401, 402, 403, 404, 413, 429 and 5xx, and never on 400, because a 400 is our own bad request.

Certificates, search visibility, crawler policy and factual claims, each swept weekly by a script that exits non-zero only on new drift. None of it is visible to a build or a test.

Stealthy backdoor triggers via dual-embedding alignment in CLIP's joint visual and textual space, reproducing and extending BadCLIP.

Targeted perturbations for multi-label classifiers that exploit semantic label co-occurrence instead of treating each label independently.

A systematic evaluation of weak-to-strong consistency (UniMatch) and self-training (ST++). The groundwork the four papers were built on.

Depth and point-cloud generation for autonomous driving on KITTI, over an improved depth-to-pointcloud pipeline on Depth Anything V2.

Why a declared context length overstates the effective one: positional-encoding bias and the STRING shift, measured across Llama, Mistral and Qwen.

Disease identification with severity scoring for smallholder farmers, plus learning modules, community data contributions and downloadable datasets.

O Level and A Level revision for African high-school students: 20+ subjects, 70+ lessons, study streaks and a learning-analytics dashboard.