Allen Gabriel M.

Chatbots rarely fail on the easy questions. They fail a few turns in, when the user rephrases, goes off-script, or the bot confidently makes something up. Finding and fixing those moments is what I do. I've evaluated 1,000+ AI responses and long multi-turn conversations, including AI agents that use tools, and I document exactly where and why a model went wrong. I also build with the Claude API in Python, so I can fix the prompt, not just report the problem. WHAT I DO • Prompt and system-prompt design: clear rules, tone, fallbacks for off-script questions, and when to refuse or hand off • Conversation flow review: dead ends, loops, wrong intent, lost context, made-up answers • Bot testing: realistic and adversarial test conversations, plus a pass/fail rubric so every change is measurable • Response rewriting: turning weak replies into ideal answers (RLHF-style) • AI safety and policy checks: harmful replies, unnecessary refusals, hallucinations, wrong-language replies • Filipino language AI: Tagalog and Hiligaynon testing, translation and native-speaker validation PROOF • Recent contract work evaluating tool-using AI agents in conversations of 100–400 turns: wrong tool use, wrong records, misreported results, rule breaks, labeled by type and severity • QA auditor on a multilingual speech-to-speech AI project: audited other raters' work on a 5-point rubric against a 72-page guideline, with evidence by turn number and exact quote • Built a Python tool that checks a project's forum through its API every day, downloads new guideline docs and flags what changed, so my work never uses outdated rules • AI data work since 2022: Alignerr (LLM evaluation and speech data), Sigma AI (Gboard Hiligaynon testing), Glimmer Tech (geospatial annotation) • 5 Anthropic certifications: Building with the Claude API, Model Context Protocol, Claude 101, AI Fluency, AI Capabilities & Limitations • 90% Job Success · 5.0 reviews: "completed the job correctly and on time" TOOLS Claude | ChatGPT | Gemini | Claude API | Python | MCP | Google Workspace HOW I WORK • I learn your bot's goal and your users first, then test against real situations • I show results with before/after tests, not opinions • I work under NDA, so I share my methods, never client data • Based in the Philippines (UTC+8), with US/EU overlap and replies within a few hours Send me a few real conversations from your bot. I'll tell you where it breaks and how I'd fix it.