Badr-Eddine – LLM, Python, RAG

A large-scale product search and discovery platform serving a catalogue of over 100 million items under hard real-time latency budgets. The technical scope spans neural retrieval, multi-stage learning-to-rank, and generative AI systems layered on top — including listing-quality classification, catalogue content generation, and cost-quality trade-off management for large models in latency-constrained pipelines.

• Owned dozens of learning-to-rank models in production carrying over $500M in annual incremental sales;
• Built an LLM system for abusive-listing detection, pairing embedding retrieval with multi-step LLM reasoning, cutting regretted impressions by 150 bps;
• Shipped generative title rewriting on AWS Bedrock (Nova) across a catalogue of over 100 million items, validated by switchback experiment;
• Productionised dual-tower neural retrieval: ONNX-exported encoders served through OpenSearch kNN and FAISS indexes, with MMR re-ranking for result diversity;
• Published "Cost-efficiency trade-offs for neural cascade rankers in web search" (ACM, 2025), on fixed-latency-budget allocation across a retrieval cascade.

An AI-powered language exam preparation platform for a low-resource European language required for citizenship naturalisation. Learners practise written and spoken answers and receive instant LLM-graded feedback against official marking criteria, exam-aligned study material, and text-to-speech pronunciation. Built and operated solo, with every scoring change gated behind a golden-answer evaluation suite before reaching users.

• Built the production scoring engine as a three-stage cascade: a deterministic rule pass, a fast-model gate, then a strong judge only where it changes the grade — most answers never reach a model;
• Made scoring provider-independent with an automated bake-off, automatic fallback to a secondary provider, and a rule-based floor that flags itself when scoring degrades;
• Built the release gate: golden-answer fixtures and CI assertions on score bands rather than exact values, so model drift doesn't break the build;
• Shipped the speech stack on GPU-backed Cloud Run using faster-whisper on CTranslate2 (int8-quantised), with round-trip character error rate tracking on the TTS side;
• Fine-tuned Whisper with LoRA on Modal and benchmarked against off-the-shelf baselines on WER;
• Built the content pipeline as retrieval-augmented authoring, with an auditor model that forces regeneration when generated material drifts from source.

An AI-powered job-search platform with paying users. Candidates describe the role they want in plain language; the platform aggregates openings from multiple job APIs, ranks them against the candidate's CV, and generates a tailored application kit per role for bulk submission. A research agent returns cited company briefings; a real-time voice agent scores mock interviews on both content and delivery. Usage data across ~200 users showed ~90% converged on the voice interview feature, and reducing friction cut landing-to-first-interview time from ~2.5 to ~2.2 minutes.

• Built the job-search agent: plain-English query, aggregation across multiple job APIs, CV-conditioned LLM re-ranking against an explicit rubric, tailored application kit per role, and unattended bulk submission
• Built a real-time voice interviewer on Gemini Live and the OpenAI Realtime API, with function calling and answers scored as the candidate speaks
• Built a research agent that plans its own report structure, issues web-search tool calls, and returns a cited PDF
• Wrote the orchestration layer without a framework, for direct control over retry policy, token spend, and latency on long-running jobs
• Ran the platform on Cloud Run with FastAPI, Firestore, and Redis, including per-call cost accounting across providers
• Owned pricing, Stripe billing, and the product analytics the team steered by

An AI image generation pipeline integrated into the admin studio of a small fashion e-commerce label. The pipeline takes a handful of reference photos of a garment and produces a complete shot set — packshots, flatlays, on-model, and detail variants — across all colourways and model categories, with a human approval step before anything reaches the storefront.

• Built the catalogue generation pipeline producing 6 shot variants per garment across 4 model categories, from a handful of reference photos, pinned to a background hex and lighting spec per set;
• Composed prompts as layered contracts — fidelity, consistency, output rules, and per-variant shot spec — with an embroidery clause pinning motif count, placement, and thread colour against the model's tendency to regularise patterns;
• Ran generation on a background worker with 30s heartbeats, 10-minute stale-job reclaim, backoff retries capped at 3, and per-model concurrency limits, swappable across gpt-image-1.5, gemini-2.5-flash-image, and gemini-3-pro-image;
• Gated publication behind human approval, with inputs, prompt, and outputs persisted per run in GCS for provenance;
• Built colour extraction on Gemini Vision: dominant garment colour isolated from background and skin, resolved to hex, snapped to the studio palette, and named in English, French, and Arabic.