Agentic AI Engineer. I build the infrastructure AI products run on.

Multi-agent systems, LLM inference pipelines, and the unglamorous production work that separates demos from products. I care about the parts that don't make it into the demo — latency budgets, cost models, error surfaces, observability — because those decide whether a product survives contact with real users.

Founding Engineer @ Browzer — CDP-native browser automation agents.

• A Chrome MV3 recorder + CDP agent with 95%+ AX/DOM element capture, cross-iframe support, and real mouse/key/upload execution.
• A streaming ReAct loop over FastAPI + extension: SSE tool execution, multi-tab orchestration, safe parallelism, abort/continue, audit logs.
• ~67% lower LLM spend via compact recording traces, context compression, prompt caching, and model routing across GPT-5 / Claude Sonnet & Haiku.
• A zero-LLM replay engine (recordings run as variable-driven tool-call templates) with a stateful AI fallback that resumes mid-run on failure.
• Self-healing docs that auto-repair on UI drift — Haiku→Sonnet diff triage, LLM-free replay of intact steps, and an agent that fixes only what changed.

~700K+ production LLM calls monitored to date.

agentflow-pro — RL for agentic reasoning. Rebuilt the ICLR 2026 AgentFlow architecture as a local Qwen3-8B Planner→Executor→Verifier→Memory loop, then trained the planner with DAPO + a learned Qwen3-0.6B Process Reward Model (replacing the paper's GRPO). Full A40 pipeline (DeepSeek-judged labels → PRM → bf16 LoRA DAPO → GGUF). +5.0 pts on GPQA-Diamond (40.0% → 45.0%) under leakage-free, quantization-matched eval.

guardloop — Guardrail runtime for async agents. Pre-flight cost/token/time/tool-call budgets, per-tool circuit breakers, a verifier-feedback retry loop, OpenTelemetry GenAI spans, and no-rewrite adapters for LangGraph & the OpenAI Agents SDK — safety enforced inside your existing LLM/tool calls.

smartmemo — Semantic memory/cache for LLM agents. Embeddings retrieve candidates; a learned pair-equivalence classifier decides reuse. Bundled classifier trained on 16,576 pairs across 9 domains — +30 precision points at equal recall vs. cosine — with implicit bad-hit detection and gated retraining.

orchflow — Dependency-free multi-agent pipelines. Typed Python for readable sequential/parallel/conditional flows: retries, shared context, flat traces, lifecycle events, human gates, and JSON checkpoint/resume.

agenteval — Behavioral eval for agents. Replaces exact-match asserts with repeated-run pass-rate tests; traces tool calls/timing/steps and emits JSON reports for CI gates. OpenAI / Anthropic / LangChain adapters.

Agents      LangGraph · OpenAI Agents SDK · LangChain · LlamaIndex · MCP/FastMCP
            ReAct · function calling · multi-agent orchestration · CDP agents
LLM eng     Fine-tuning (LoRA/QLoRA/PEFT) · DPO/GRPO/DAPO/PRM · vLLM/SGLang/Ollama
            quantization (AWQ/GPTQ/GGUF) · context engineering · KV-cache · streaming
Retrieval   PyTorch · sentence-transformers · RAG · FAISS · Qdrant · pgvector
Eval/Obs    OpenTelemetry · Langfuse · Arize Phoenix · LangSmith · GPQA/AIME
Systems     FastAPI · Nest.js · GraphQL · SSE/WebSockets · Docker · GCP · AWS · Redis
Languages   Python · TypeScript · C++ · SQL

abhinandan.one · linkedin/in/abhibuilds · hi@abhinandan.one

_{Open-source contributor at HeroUI (YC S24) — fixed 10+ bugs and shipped 7+ enhancements across Calendar, Table, and Pagination, which led to a personal offer from the CEO.}