🧠 AI Engineer based in Bengaluru, India | 8+ years in AI/ML, last 3 years deep in agentic systems, evals & post-training

I'm an AI engineer with 8+ years of experience across AI and machine learning — the last 3 years focused specifically on agentic systems, LLM evaluation, and model training. I design the environments, harnesses, and reward signals that make agents measurably better, including building RL environments for evaluating coding agents on SWE-bench and SWE-bench Pro. I've also founded a startup, shipped production ML at fleet scale, and led fine-tuning + eval work on multi-agent legal AI products.

• 🏗️ Design & ship multi-agent systems in LangGraph — supervisor/subagent orchestration, plan-execute-replan control flow, and dynamic tool loading
• ⚖️ Led engineering on Sidebar, a multi-agent legal AI product — production system with real revenue and measurable efficiency gains for users within months of launch
• 🔬 Empirically compared supervisor, mesh, and ReAct agent topologies to pick the right architecture for the task, not just the trendy one
• 🔀 Built a multi-model routing layer (LiteLLM) across GPT-4, Claude, and Gemini with production fallbacks — optimizing the cost/latency/accuracy frontier
• 🛠️ Tool use & function calling against live APIs and external databases in production, not just demos

• 🎮 Built RL environments for evaluating coding agents — task instantiation and reward signal for benchmarks like SWE-bench and SWE-bench Pro
• 🧱 Designed a three-tier agentic evaluation harness: per-subagent checks, inter-agent handoff fidelity, and end-to-end outcome scoring
• 📼 Built structured execution traces — replayable, step-attributable, diffable across model/prompt versions (think: stack traces, but for agents)
• 🎯 Rubric-based & LLM-as-judge grading, golden-dataset construction, plus objective retrieval metrics (recall@k, nDCG) where ground truth exists
• 🏷️ Trained verifier-style classifiers on LLM-generated labels (weak supervision / label distillation) when human labels didn't scale
• 🌱 Currently exploring open-source RL environments in the legal domain — deterministic task instantiation, programmatic verifiers, reward at every decision point

• 🚀 Supervised fine-tuning of Qwen3 8B for structured extraction — beat a prompted frontier-model baseline on cost and latency
• 🧬 Fine-tuned embedding models for domain-specific retrieval
• 🗜️ Distillation & quantization for real-time inference on 1M+ edge devices
• 🎨 GAN-based synthetic data generation for domain adaptation
• 🕸️ Multi-task deep learning with shared-backbone architectures (detection + tracking + distance estimation, jointly optimized)

• 🔎 Hybrid search (BM25/BM42 + dense embeddings), learned reranking, knowledge-graph-augmented RAG
• 🏭 Fault-tolerant document pipelines on Apache Beam / Google Cloud Dataflow at scale
• 🗄️ Vector DBs: Pinecone, Qdrant, Turbopuffer, FAISS, Milvus, Weaviate, Elasticsearch

Python C++ Go TypeScript SQL LangGraph LangChain LiteLLM LlamaIndex PyTorch TensorFlow Hugging Face Apache Beam Dataflow Docker Kubernetes GCP Azure FastAPI LangSmith Langfuse

• 🚀 Co-Founder & CTO of Stybe, an AI-powered fashion e-commerce startup — built the full RAG-based personal-stylist stack from scratch
• 🚛 Built perception ML at fleet scale at Motive — multi-task detection/tracking models optimized and quantized for 1M+ edge devices (dashcams), plus a facial-recognition (Driver ID) pipeline running at scale
• 📄 2x Springer publications on pedestrian-AV interaction and safety
• 🎓 M.Tech in Computer Engineering & AI, College of Engineering Pune

📫 Reach me: LinkedIn · pranavpandey2511@gmail.com