Projects

Production AI systems, research, and applied ML spanning NLP, low-resource languages, and AI engineering.


AI Engineering Portfolio

20 production-grade AI systems. Every project is fully runnable end-to-end.  View on GitHub →

01 — Production RAG PDF Q&A with page citations. Hybrid search (dense + BM25), cross-encoder reranking, LangGraph pipeline. ::: {.tech-stack} LangGraph SQLite-vec BM25 FastAPI

02 — Model Router Routes queries to Haiku, Sonnet, or Opus by complexity. Saves up to 10× on API cost. ::: {.tech-stack} Anthropic SDK Prometheus FastAPI

:::

03 — Multi-Agent Research CrewAI pipeline: Researcher, Writer, Fact-Checker with human-approval gate and JSONL audit trail. ::: {.tech-stack} CrewAI Anthropic SDK Audit Trail

:::

04 — Eval Harness 20+ golden Q&A test cases run before every deploy. Blocks CI if quality drops below threshold. ::: {.tech-stack} DeepEval RAGAS pytest

:::

05 — Observability Full-stack AI observability: OpenTelemetry traces, Prometheus metrics, auto-provisioned Grafana dashboard. ::: {.tech-stack} OpenTelemetry Prometheus Grafana

:::

06 — Security Guardrails Middleware: prompt injection detection, PII redaction with Presidio, output filtering, rate limiting. ::: {.tech-stack} Presidio FastAPI slowapi

:::

07 — Local Dev Environment Zero API cost RAG stack that runs 100% offline. Ollama + LanceDB + sentence-transformers. ::: {.tech-stack} Ollama LanceDB FastAPI

:::

08 — Streaming Copilot UI Token-level streaming chat with Next.js 14 App Router and Vercel AI SDK. ::: {.tech-stack} Next.js 14 Vercel AI SDK TypeScript

:::

09 — Fine-Tuning Pipeline LoRA fine-tuning with PEFT + DPO, ROUGE-L evaluation, and catastrophic forgetting detection. ::: {.tech-stack} PEFT TRL Transformers

:::

10 — Multi-Tenant SaaS JWT auth, row-level tenant isolation, per-tier rate limiting, usage cost accounting. ::: {.tech-stack} FastAPI SQLAlchemy JWT

:::

11 — CI/CD for AI GitHub Actions eval gate → ArgoCD canary deploy → smoke tests → auto-rollback. ::: {.tech-stack} GitHub Actions ArgoCD Kubernetes

:::

12 — Vector DB at Scale Qdrant hybrid search with Reciprocal Rank Fusion and Redis embedding cache (24h TTL). ::: {.tech-stack} Qdrant Redis FastAPI

:::

13 — Agent Memory Redis short-term buffer + Qdrant long-term recall + automatic context compression. ::: {.tech-stack} Redis Qdrant Anthropic SDK

:::

14 — Inference Load Balancer vLLM proxy with health checking, least-queue-depth routing, TTFT metrics, auto-failover. ::: {.tech-stack} vLLM Prometheus Kubernetes

:::

15 — Human-in-the-Loop Pauses AI on irreversible actions for human review via a WebSocket approval UI. ::: {.tech-stack} LangGraph WebSocket FastAPI

:::

16 — Agentic Automation Async job queue with HMAC webhook validation, Redis idempotency, Celery exponential backoff. ::: {.tech-stack} Celery Redis FastAPI

:::

17 — Domain Benchmark Legal AI evaluation suite. ROUGE-L + semantic similarity scoring with a public leaderboard. ::: {.tech-stack} DeepEval ROUGE sentence-transformers

:::

18 — Sunflower Translator Translation across 31 Ugandan languages via Sunbird Sunflower-14B-W8A8. ::: {.tech-stack} Sunflower-14B vLLM FastAPI

:::

19 — Uganda Health Bot Community health chatbot in Luganda, Acholi, and English with emergency detection and WhatsApp. ::: {.tech-stack} Sunflower-14B FastAPI Twilio

:::

20 — Model Drift Monitor Probes a model every 5 minutes, detects semantic, behavioral, and latency drift. Warning/critical alerts. ::: {.tech-stack} sentence-transformers scipy SQLite

:::

:::


Research & Earlier Projects

Text Classification Using Embeddings for Luganda

Topic classifier for the Luganda text corpus using Cohere embedding representations and an SVM with hyper-parameter search. Demonstrates full NLP pipeline development for a low-resource Bantu language.

Python SVM Cohere API Scikit-learn Low-resource NLP

Water Quality Data Visualisation — Kampala

Interactive visualisation of water quality metrics across Kampala Central Division, surfacing spatial patterns in sanitation data to support public health analysis and policy.

R Data Visualisation Geospatial Analysis Public Health

Fine-Tune Alpaca for Any Language

Fine-tuned the Alpaca instruction-following LLM for multilingual language translation and topic classification, exploring parameter-efficient adaptation for low-resource language settings.

Python LLaMA / Alpaca Instruction Tuning Fine-tuning Multilingual NLP

Chest Opacity Detection from Ultrasound Images

Deep learning pipeline for detecting chest opacities in ultrasound images — a low-cost, non-ionising alternative to CT scanning, aimed at resource-constrained clinical settings in LMICs.

Python PyTorch Deep Learning Medical Imaging Computer Vision