Production AI systems · Agentic AI · RAG · MLOps

Two systems, built to ship.

Two production-grade AI systems, built end-to-end — data pipelines, hybrid & graph RAG, post-training, CI-gated evals, and scale-to-zero deployment. Each ships with a live demo: deterministic where it counts, grounded in real data, and running live end to end.

Agentic AI Hybrid & Graph RAG Fine-tuning & inference MLOps CI-gated evals
01

TaxBrainAI

Agentic Indian income-tax advisor

Computes deterministically to the rupee, grounds every answer in the statute with an eight-stage hybrid & graph retrieval pipeline, and runs a fail-closed verifier that refuses rather than guess. Built on a nine-node agent graph with human-in-the-loop approval; the language model transcribes and cites — the numbers come from a property-tested engine, never the model.

0.99 faithfulness 100% exact to the rupee 0 PII leaks 366-row held-out eval
LangGraph Hybrid + Graph RAG Fine-tuned Gemma 4 E4B Deterministic engine FastAPI · SSE MCP · 8 tools vLLM · scale-to-zero
Open the demo →
02

MLPipeline Pro

Real-time payment-fraud MLOps

A production MLOps reference for payment-fraud detection — gradient-boosted scoring with drift detection, SHAP explainability, and a robustness-gated CI pipeline, served through a reproducible train-to-deploy lifecycle on a scale-to-zero endpoint.

ROC-AUC 0.915 PR-AUC 0.552 segment-robust 0.88–0.95 116 tests
LightGBM Dagster MLflow Feast Evidently SHAP · Fairlearn BentoML · Modal
Open the demo →
02

Skills & stack

LLMs & Agentic AI

LangGraphMCP · FastMCPA2A protocolClaudeGPT GeminiGemma 4 E4BPydanticAWS Bedrock Agents Anthropic SDKLiteLLMtool / function callingprompt engineering

Fine-Tuning & Inference

SFTLoRADPOGRPOGSPORL PyTorchHF TransformersAccelerateUnslothTRL vLLMFP8AWQ-INT4

RAG, Retrieval & LLM Safety

RAGhybrid retrievalGraph RAGHyDEBGE-M3 GraphitiLLMLingua-2multimodal visionstructured output PresidioPrompt Guard 2Llama GuardSpotlighting

ML, MLOps & Evaluation

XGBoostLightGBMOptunaMLflowW&B DagsterFeastEvidentlyNannyMLBentoML SHAPFairlearnDeepEvalInspect AIPromptfoo ArgillaLLM-as-judge · G-EvalLangfuse

Backend & APIs

PythonJavaTypeScriptFastAPISpring Boot Spring SecurityREST · SSEMicroservicesOAuth2 / OIDC · JWT JPA / HibernateResilience4jSQLAlchemyKong API GatewayNext.js

Data & Databases

PostgreSQLMySQLRedisClickHouseNeo4j pgvectorSQLApplied Statistics

Cloud, DevOps & Observability

AWSGCPKubernetesDockerTerraform · Helm ModalCloudflare WorkersEC2 · S3 · SQSBedrock Pub/Sub · DataflowGitHub ActionsArgoCD · Jenkins Prometheus · GrafanaELKOpenTelemetry

Engineering Practices

Distributed SystemsSystem Design (HLD/LLD)ML System Design Event-Driven ArchitectureREST API DesignTDDJUnit 5 · Mockito Property Testing (Hypothesis)Performance Optimization
03

Open source

Published · PyPI · MIT

taxbrainai-compute

The deterministic Indian income-tax engine behind TaxBrainAI, as a standalone open-source library. Computes both the new and old regime for AY 2026‑27 — slabs, the Section 87A rebate with marginal relief, surcharge caps and marginal relief, capital gains (Section 112 indexed lower-of and the 23 Jul 2024 cut-over), crypto / VDA (Section 115BBH), AMT, and statutory Section 288A/B rounding — with every rate cited to its source section. Pure-Decimal · 139 tests at 100% coverage (20 property-based invariants × ~10K generated scenarios via Hypothesis) · cross-verified against all four Income-Tax Department e-filing utilities. Zero dependencies, MIT.

PyPI → GitHub →
Published · MCP Registry

taxbrainai-mcp

A FastMCP server that exposes the tax engine as 8 tools over the Model Context Protocol — plug deterministic Indian income-tax computation and statute lookup into any MCP-compatible client. Listed on the official MCP Registry as io.github.harshil-projects/taxbrainai-mcp.

PyPI → GitHub → MCP Registry →
Published · Hugging Face

taxbrainai-gemma-4-e4b

The advisor-synthesis model behind TaxBrainAI — Gemma 4 E4B, fine-tuned to transcribe the engine's figures and cite the governing statute, never to author a number. LoRA adapters + merged weights, released on the Hugging Face Hub under Apache-2.0.

Hugging Face →

The builder

Want this kind of engineering on your team?

Both systems are my work end to end — architecture, data, training, evals, deployment, and these pages. I’m open to agentic-AI, ML-platform, and backend roles.

Email me GitHub