Anmol Jaiswal anmolg1997

Contributor to Google ADK · Pydantic AI · LiteLLM · llama.cpp · ag-ui
_{7+ years in applied AI · author of a package in the official ADK integrations catalog}

I build production AI systems that reason, plan, and execute autonomously: multi-agent orchestration, agent skills and harnesses for coding agents, enterprise RAG pipelines, LoRA fine-tuning at scale, and multi-adapter inference serving.

Focus area	Working with
Multi-agent AI	`Google ADK` `A2A Protocol` `MCP` `Agent Orchestration`
Agent skills & harnesses	`SKILL.md` `Claude Code` `Operator Skills` `Evidence Gates`
Knowledge graphs & RAG	`Neo4j GraphRAG` `Hybrid Search` `Reranking` `Guardrails`
LLM fine-tuning & serving	`LoRA / QLoRA` `Unsloth` `vLLM` `Multi-Adapter Inference`
LLM observability	`Langfuse` `MLflow` `OpenSearch` `Domain Evaluation`
Model building	`Transformers from scratch` `DeepSpeed` `SFT / DPO / RLHF`

In industry

Building an enterprise decision-intelligence platform: multi-agent orchestration on Google ADK, NL2SQL over governed data, AG-UI streaming to a React frontend
Production hardening: durable sessions, tenant isolation, model benchmarking, incident postmortems

Now building

Project	What it is
rhytm	Multi-agent framework for proprietary data analysis with business intelligence
SLM-From-Scratch	Ongoing: alignment utilities and GGUF / ONNX export paths for small language models

Tracking the frontier

Where agentic AI is heading in 2026, and where I'm already hands-on:

Frontier	My work there
Agent skills: portable `SKILL.md` capabilities for coding agents	fincept-operator-skill: GUI operation of a live trading terminal with human-approval gates prepostmortem-skills: failure analysis shipped as skills
Harness engineering: the control plane around coding agents	harness-engineer-skill: audits any repo's harness, any convention Runs the Fresh Session Test, enforces an evidence-before-done gate
Agent protocols: `MCP` `A2A` `AG-UI` `AGENTS.md`	ADK 2.0 middleware for ag-ui A2A orchestration in Multi-Agent-AI-Framework
Context engineering: sessions, memory and state over prompts	adk-database-memory: durable agent memory Firestore session service patterns adopted upstream in Google ADK
Agent reliability: evals, failure taxonomies, observability	261-incident, 12-class failure model in prepostmortem-skills Retry-safe LLM-as-judge eval merged into Pydantic AI

AI x markets

EquityOracle: self-improving equity recommender with multi-horizon predictions and paper trading
fincept-operator-skill: agent skill operating a live trading terminal, paper-only with human-approved live actions

The journey

Years	Era	Proof
2019–2021	Data science and classical ML	Lead-Scoring · Telecom-Churn · Product-Recommender
2021–2023	NLP and deep learning	Deep-Learning-Projects · Spark-GPU-Sentiment
2023–2025	GenAI foundations: RAG, fine-tuning, model building	Enterprise-RAG · LLM-Finetuning-Toolkit · SLM-From-Scratch
2025–now	Agentic AI: orchestration, skills, harnesses	Multi-Agent-Framework · adk-database-memory · prepostmortem-skills

Google ADK Python	First Firestore session service: transactional state, subcollection events, batch deletes Patterns adopted in the official implementation, credited by a Google engineer
ADK Community	`FirestoreSessionService` with 19 unit tests and in-memory mocks Race-safe transactions, N+1 query elimination, async batch deletes
ag-ui Protocol	ADK middleware upgraded to `google-adk 2.0`: +975/−48 LOC, closed 2 upstream issues LangGraph adapter fixes: routes, fork config, stream message IDs
Pydantic AI	Merged: retry-safe LLM-as-judge evals via constrained reason field (#5089) History processor repairing orphaned tool calls behind provider 400s
llama.cpp	Type + integer-range validation in `GGUFWriter.add_key_value` Catches overflow before it silently corrupts model metadata
LiteLLM	Merged: KeyError crash fix in Anthropic file-id discovery (#26228) Deterministic aiohttp session disposal, fixing leaks (#32003)
adk-database-memory	SQL-backed ADK memory: Postgres / MySQL / SQLite, drop-in for `InMemoryMemoryService` Listed in the official ADK integrations catalog
(pre·post)mortem	261 tagged real-world incidents, 12-class agentic-AI failure model Shipped as agent skills for pre- and post-change risk review

SLM From Scratch

BPE tokenizer + composable Transformer: RoPE, GQA, SwiGLU
DeepSpeed training, SFT / DPO / RLHF alignment, GGUF / ONNX export

PyTorch DeepSpeed RLHF

Enterprise RAG System

Hybrid search (dense + BM25) with cross-encoder reranking
Guardrails, semantic caching, RAGAS evals, Langfuse observability

LlamaIndex OpenSearch Langfuse

Multi-Agent AI Framework

Coordinator, planner, coder and reviewer agents on Google ADK
YAML-driven personas, A2A protocol

Google ADK A2A FastAPI

NL2SQL Engine

Live schema introspection, self-correction, few-shot learning
Multi-dialect: Postgres, Snowflake, SQLite

SQLGlot PostgreSQL Snowflake

LLM Fine-tuning Toolkit

Pluggable backends (Unsloth, TRL, Axolotl) driven by YAML recipes
MLflow tracking, vLLM serving

Unsloth vLLM MLflow

Multi-LoRA Serve

One base model, many LoRA adapters per request
OpenAI-compatible gateway, tenant routing, Prometheus metrics

vLLM LoRA Prometheus

KG-RAG

Neo4j knowledge graph + RAG for document QA
Entity extraction, graph-traversal retrieval, hybrid grounding

Neo4j FastAPI React

Domain-Adaptive LLM

Medical, legal, finance and code specialization via curriculum training
Domain benchmarks (MedQA, LegalBench), safety guardrails

MedQA LegalBench LoRA

Full tech breakdown

LLM providers      OpenAI · Anthropic · Google Gemini · Llama · Mistral
Agent frameworks   Google ADK · A2A Protocol · MCP · LangGraph · CrewAI
RAG stack          LlamaIndex · LangChain · Neo4j · OpenSearch
Vector DBs         Pinecone · Weaviate · Milvus · Qdrant · ChromaDB
Observability      Langfuse · MLflow · Weights & Biases · OpenTelemetry
Inference          vLLM · Multi-LoRA serving · TensorRT-LLM · ONNX Runtime
Fine-tuning        LoRA · QLoRA · DoRA · Unsloth · Axolotl · DeepSpeed · RLHF / DPO
Model building     PyTorch Transformers · BPE tokenizers · GGUF / ONNX export
Frontend           React · Vite · Next.js · TypeScript · TailwindCSS
Backend            FastAPI · Python · Node.js · GraphQL
Cloud              AWS (Bedrock, SageMaker) · GCP (Vertex AI) · Azure (OpenAI, AI Search)
Data & distributed PySpark · Spark NLP · Databricks · Streamlit · scikit-learn · XGBoost
Infrastructure     Docker · Kubernetes · Terraform · GitHub Actions

GitHub stats GitHub streak

Top languages

Building something complex? Let's talk.

Navigation Menu