João – Python, RAG, PostgreSQL

A multi-tenant SaaS platform for biotech and research labs, centralising inventory, samples, equipment, calibrations, maintenance, and SOPs into one governed system with per-organisation data isolation, role-based permissions, and full audit trails. Built for regulated lab environments where clean, queryable, permissioned data is the precondition for any further automation or AI features.

• Architected and shipped the entire platform solo — data model, backend, UI, integrations, and deployment — from concept to production.
• Designed multi-tenant data architecture with strict per-organisation isolation, query scoping, and role-based access control across lab and org roles.
• Modelled complex scientific-domain entities (inventory lots, locations, samples, equipment, calibrations, SOPs, audit records) into a clean, queryable schema.
• Built an SOP lifecycle engine (draft → review → approval → active → superseded) with version acknowledgements and full change traceability via audit logging.
• Engineered data pipelines: CSV import/export with validation and preview, a versioned REST API for external system integrations, and scheduled email/in-app notification digests for expiry, stock, and calibration alerts.
• Integrated Stripe billing and containerised the full stack with Docker Compose for repeatable deployment.

A production SaaS platform for Portuguese freelancers working with international clients. Handles the full financial layer: foreign-currency income, platform-fee reconciliation, VAT, Social Security, and IRS planning — tax-deadline tracking, financial projections, and reporting, all in one place.

• Designed, built, and shipped the product end-to-end as sole founder and engineer.
• Built financial data ingestion and transaction-management workflows with multi-currency conversion and platform-fee handling.
• Implemented VAT, Social Security, and IRS planning calculations alongside tax-deadline tracking and financial projections.
• Developed reporting and data export functionality for freelancer tax compliance.
• Integrated external APIs and third-party services into the platform.
• Deployed the full stack (Python, FastAPI, PostgreSQL, React, TypeScript) via Docker on a VPS.

A multi-tenant data platform for a German mystery-shopping agency, serving ~30 enterprise clients across retail and automotive, each with configurable dashboards and AI-generated analytics reports. At the core: an LLM-powered report-generation system that turns raw evaluation data into client-ready analytical reports using structured (schema-validated) OpenAI outputs and per-client configurable prompts — AI as a production feature across every tenant, not a one-off integration.

• Designed and shipped an LLM-powered monthly report generator: aggregated mystery-shopping evaluation data → OpenAI Responses API with Pydantic schema validation → typed reports (summary, strengths, improvement areas, prioritised actions) → branded German PDF.
• Built per-tenant AI configuration — each client's system prompt is a DB record editable from admin, so report tone, domain guidance, and focus are tuned per client without code changes.
• Production-hardened the AI pipeline: rate-limit and quota handling, automated email alerting on pipeline failures, and graceful user-facing fallbacks.
• Architected multi-tenant structure as one Django app per client, scaffolded from a config-driven generator for repeatable, low-risk onboarding.
• Engineered read-only data-ingestion pipelines from Google Sheets into PostgreSQL with row-level validation and scheduled syncs via a task queue (procrastinate), keeping ingestion off request handlers.
• Integrated Google Cloud Translation for free-text fields, SSO (Microsoft Entra ID + Google), and deployed on AWS Elastic Beanstalk with RDS PostgreSQL.

A distributed, containerised AI news aggregation and publishing platform that ingests news at scale, classifies and enriches articles with AI, rewrites and summarises content using LLMs, and publishes it automatically. Built as multiple orchestrated services — scraper, zero-shot GPU classification microservice, self-hosted LLM server, and Django web app — deployed via Docker Swarm with a shared model layer.

• Designed a provider-agnostic LLM orchestration layer supporting OpenAI (chat, vision, reasoning, DALL·E), Anthropic (including extended thinking), xAI Grok, Perplexity, a self-hosted llama.cpp server, and local Stable Diffusion — all behind one uniform calling interface.
• Built a runtime-configurable model registry: each model (provider, endpoint, credentials, calling function) is a DB record, so which LLM runs each pipeline step is swappable from admin with no redeploy.
• Built a self-hosted, GPU-accelerated zero-shot classification microservice (HuggingFace Transformers, CUDA/PyTorch, threshold-based multi-label) exposed as a Flask service; paired with an LLM-based two-tier category assigner triggered via Django signals.
• Integrated a containerised llama.cpp inference server (GPU Dockerfile) with a custom LangChain LLM wrapper for local, cost-controlled generation and RAG alongside hosted APIs.
• Instrumented the full pipeline with Langfuse observability (per-workflow traces on every generation) and tenacity exponential-backoff retries for resilient production LLM calls.
• Implemented dynamic prompt management with Jinja2-templated system/user prompts and runtime data injection, plus DB-stored prompt sets per feed and per workflow.
• Deployed the multi-service system via Docker Swarm, sharing an installable internal models package across services.

An AI literary-analysis assistant built for a publishing editorial team. Allows users to upload books and chat with a tool-calling agent that retrieves relevant passages, answers questions, and compares content across titles — all grounded in the source text with inline, verifiable citations. Features hybrid keyword + semantic retrieval with rank fusion, an anti-hallucination citation contract, and a dedicated evaluation harness for retrieval accuracy and answer faithfulness.

• Designed a tool-calling agent with a bounded 5-round multi-turn loop orchestrating five tools (search, retrieve, compare, summarise, cross-reference) for grounded literary Q&A and cross-book analysis.
• Implemented hybrid retrieval combining SQLite FTS5 BM25 and sqlite-vec cosine KNN, fused with Reciprocal Rank Fusion, over paragraph-aligned, chapter-bounded chunks (~600 tokens, ~15% overlap).
• Engineered a citation-grounding and anti-hallucination contract: tracked chunk IDs from tool calls, stripped any citation the model invented, and linked every UI reference to a verifiable passage.
• Built a two-mode evaluation harness: retrieval hit-rate@k and an end-to-end faithfulness check that parsed inline citations from the live answer stream and verified each cited chunk against source text.
• Implemented crash-safe background ingestion (HTML parse → chunk → embed → LLM summaries) with a durable queue that re-scheduled interrupted work on restart, keeping chat and search fully unblocked.
• Integrated Azure OpenAI (chat + embeddings) with exponential-backoff resilience and SSE token streaming; delivered full-stack: React + Vite frontend with streaming chat, book manager, and outlines.