Lalit Kushwaha

Audited complex multimedia generations for temporal consistency, using deep acoustic artifact detection and executing high-precision bounding-box and video lip-sync verification. Evaluated chain-of-thought logic and ensured strict instruction adherence including JSON/regex formatting validation, rigorous answer-key verification, and native Devanagari orthography QA. Designed adversarial red-team prompts to stress-test safety guardrails and systematically audited outputs for toxicity, bias, and factual hallucinations against verified databases. Cataloged high-volume evaluation datasets via Google Sheets and enterprise annotation platforms, maintaining strong inter-rater reliability while exceeding throughput targets. Also assessed multimodal audio dubbing and AI-dubbed media for prosody, intonation, and translation fidelity to ensure naturalness and acoustic consistency; verified transcription accuracy and applied spatial annotation QA (bounding boxes) plus video lip-sync and temporal c

Assessed LLM responses and search index results against E-E-A-T and YMYL frameworks, grading semantic intent, factual accuracy, and safety compliance. Evaluated synthesized cross-lingual audio and AI-dubbed media for prosody, intonation, and translation fidelity to ensure native-level naturalness and acoustic consistency. Performed stringent Devanagari/English transcription and grammar review for contextual accuracy alongside spatial bounding-box annotation QA and video lip-sync/temporal consistency verification. Navigated detailed rating rubrics, documented data set anomalies, ensured alignment with evolving project SLAs, and reduced downstream human editing time through semantic deduplication and strong QA rubric execution.

Engineered structured system prompts, few-shot question blueprints, and multi-step agentic workflows to automate generation of complex, syllabus-aligned diagnostic datasets (SSC and Stenographer benchmarks). Analyzed historical examination patterns to map topic distributions and difficulty tiers, setting source-grounded constraints to guide LLM trajectory learning and prevent semantic drift. Audited step-by-step explanations and answer keys against verified ground-truth documents, removing logical ambiguity, reasoning flaws, and non-compliant patterns across bilingual (Hindi/English) outputs. Developed end-to-end quality assurance rubrics for difficult calibration, formatting compliance, and repeated-content (semantic deduplication) detection, reducing downstream human editing time by 80%.