Hi, I'm Aniket SakpalI build AI Agents that
reason, create & take action

I design the reasoning loops, memory architectures, and tool systems that make LLM agents reliable in production — at Expedia Group and through my own startups.

• →Built production-grade AI analytics copilot with adaptive reasoning loop — ReAct tool selection, multimodal execution, and episodic history compaction.
  • →Declarative skill system — YAML-manifest-driven workflows (metric diagnosis, root cause analysis) with hot-loading to eliminate LLM drift.
  • →Redis state management — key-sharding, TTL policies, partial-state hydration for sub-second agent transitions.
  • →Tiered memory architecture — factual, episodic, and preference memories with vector compression, relevance decay, and attention-gated retrieval.
  • →Context-engineering layer — dynamic pruning, hierarchical chunk selection, and semantic routing to optimize LLM context windows.
  • →Knowledge graph RAG — MBR/WBR retrieval via KG expansion, embedding debiasing, query rewriting, and MMR reranking.
  • →Production orchestration — async FastAPI + Redis Streams on Kubernetes; Datadog APM + Langfuse LLM tracing.
• →Built patented agentic image-to-video system — first in travel at scale, delivering $7M+ marketing uplift across 10,000+ hotel properties.
  • →Hallucination detection — Vision-LLM reasoning, object-mask tracking, and temporal cross-attention validation.
  • →Camera-motion detector — RAFT-style optical flow + custom video-embedding encoders to flag jitter and misaligned trajectories.
  • →Autonomous regeneration loop — LLM-driven prompt optimization, shot-plan adjustments, and diffusion-model control tuning.
  • →Image-selection module — fine-tuned YOLOv8 + CLIP dual-encoder on 1M+ images; LangGraph workflow for frame scoring and quality gating.
  • →Video upscalers — 3D-UNet + latent-space SR transformers with GAN-based perceptual loss and motion-aware temporal consistency.

• →Built AI learning co-pilot transforming documents into interactive notebooks — explainer videos, Socratic audio, adaptive quizzes, and text Q&A.
  • →Multi-agent architecture — RAG, code generation, visualization, evaluation, and TTS agents coordinated for structured learning modules.
  • →RL fine-tuning (GRPO/RLHF) and multi-agent orchestration powering scalable content-assembly pipelines.

• →Developed Search & Ranking Interpretation Framework — latent-class discrete choice models, causal inference, and ranking-explainability diagnostics.
  • →Latent-class choice models via custom EM algorithm — uncovered heterogeneous segments, utility functions, and substitution patterns driving conversion.
  • →Structural causal modeling — instrumental variables, propensity weighting, and counterfactual estimators to isolate true booking drivers from bias.
  • →Ranking-evaluation layer — NDCG, ERR, Shapley/IG attributions to diagnose where algorithmic ordering diverged from customer utility.
• →Owned A/B testing infrastructure for Ranking ML models — statistically rigorous, interpretable evaluation at scale.
  • →Interleaving pipelines (team draft / probabilistic) enabling 10× faster model iteration with variance-reduced comparisons.
  • →Interpretable ranking metrics — qualified CTR, position-normalized engagement, and intent-aligned utility gains.

• →Improved app rating 3.1 → 3.7 leading product analytics and UX experimentation — drove six major feature launches; built funnel-conversion, marketing-effectiveness, and KPI alerting dashboards.

• →Led team of 5 delivering ML for Fortune 100 CPG and Fortune 1 retailer — predictive modeling, optimization, and large-scale analytics.
• →Reduced churn 11% → 8% — survival models (Cox PH, Weibull AFT) + clustering (k-means/GMM) to identify high-risk segments.
• →Improved lead conversion 26% → 29% — regularized logistic-regression scoring and uplift-based ranking of sales prospects.
• →Increased manufacturing uptime 86% → 92% — gradient boosting failure detection and real-time scoring pipelines.
• →Marketing mix modeling (MMM/ROI) and market-basket analysis via association-rule mining for strategic product placement.