I evaluate and improve AI systems for a living: I review how models generate code, catch where they hallucinate or break, and build the rubrics and verifiers that keep output quality honest. I write Python daily and reach for Java, C, Go, or TypeScript when the task calls for it. I recently reviewed frontier AI at Scale AI (Outlier) and I am now open to new roles.

My focus: find the exact turn where a model breaks, and build the deterministic check that proves it.

Three public tools that show one idea - evaluation as a hard gate, not a vibe - from three angles.

• AI Trainer and Reviewer - Scale AI (Outlier), 2026. Promoted from Attempter to Reviewer. Golden/silver agentic RL trajectory pipelines, rubric and pytest-verifier design, and safety-taxonomy grading (OpenClaw Atlas, Blue Shell, Lobster Safety). Confidential work, shown as process only.
• Software and ML Engineer - Tensium, 2026 to present (remote). CPU-based long-horizon ML tasks with multistep reasoning and code generation; develop, debug, and refactor Python pipelines for maintainability and performance.
• AI Trainer, LLM Evaluation - Handshake AI, 2026. Adversarial multi-hop question authoring plus an automated QC checker (Project Seal).
• MCX Advisor - Zebu Share and Wealth Management, 2025 to 2026. 4 years of trading experience across crypto and Indian equities; supported client trading operations and profit-and-loss updates.

Public process walkthroughs of the evaluation work (nothing confidential): pspmani.github.io/workflows.html

• Promoted from Attempter to Reviewer within 3 months at Scale AI (Outlier), for evaluation accuracy and consistent task quality.
• Shipped three live, public AI-evaluation tools: TrajLens, JudgeLab, and EvalGate.
• EvalGate CI/CD pipeline has run green for months, with automated weekly retraining and quality-gated deployments.
• Caught a session-level data-leakage bug and corrected an inflated 99% accuracy to a truthful 92.5% using GroupKFold.
• Vice President, Iterators Club - Jansons Institute of Technology (student leadership during B.E.).

• HandPilot - webcam gesture control for Windows: MediaPipe and OpenCV hand tracking for pointer, pinch clicks, scrolling, and an on-screen keyboard, with Windows CI tests.
• Chatbox - streaming chat app: FastAPI and Server-Sent Events backend, React and TypeScript frontend, pluggable model backend that runs offline with no API key.
• Customer Churn Predictor - Gradient Boosting on 7,043 records, 0.843 test ROC-AUC, SHAP explanations, deployed on Streamlit.
• Diabetes Foot Pressure Classification - caught session leakage with GroupKFold, correcting 99% to a realistic 92.5%.
• Spam Email Classifier - NLTK and TF-IDF pipeline, 97.8% accuracy with a Linear SVM.

• Claude Code 101 - Anthropic (2026)
• AI Fluency for Builders - Anthropic (2026)
• Introduction to Agents - Mercor (2026)
• Prompt Engineering (OutlierEDU) - Outlier AI (2026)
• Problem Solving (Intermediate) - HackerRank
• Frontend Developer (React) - HackerRank
• AWS Academy Cloud Foundations - Amazon Web Services (2022)
• Microsoft 365 Productivity Advanced - Microsoft (2022)
• Ethical Hacking and Penetration Testing - Scholiverse
• Enrolled Agent (EA) credential - in progress

Evaluation: LLM evaluation, RLHF / RLAIF, red teaming, prompt engineering, rubric design, agentic trajectory review, hallucination detection, multimodal assessment.
Also: US Enrolled Agent (EA) credential in progress; Indian income-tax returns (ITR-1 to ITR-4).

Open to AI Evaluation, LLM Engineering, RLHF / RLAIF, and ML Engineer roles.
Portfolio  |  LinkedIn  |  Email