Atharva Joshi · Forward Deployed Engineer

01 · The Human

About01

I'm a Forward Deployed Engineer who builds generative-AI systems that survive contact with production. My focus is the hard last mile: taking LLM agents, RAG pipelines, and agentic workflows from prototype to reliable, governed, real-world deployment. I embed with the problem, map chaotic enterprise data into something a model can actually reason over, and ship systems that stakeholders trust, backed by evaluation, guardrails, and audit trails. Recent work spans healthcare AI, clinical-data harmonization, and open-source contributions across the Microsoft, Google, and Goldman Sachs ecosystems.

// Always Supporting

Man United

Real Madrid

India Cricket

Buffalo Bills

// Over the Board

♞ chess.com / aaichamulga ↗ lichess / aaichamulga ↗

atharva@portfolio:~

// Right Now

Reading

Me Before You

Jojo Moyes

Listening

The Night We Met

Lord Huron

▶ NOW PLAYING

Watching

Grave of the Fireflies

Isao Takahata · 1988

// FUEL

Cristiano Ronaldo

Football

"Hard work beats talent"

Novak Djokovic

Tennis

"Believe in yourself"

Kobe Bryant

Basketball

"The details are not the details"

Bobby Fischer

Chess

"Chess is life"

// Local Time · New York

02 · Selected Works

Projects That Ship02

~/eligible · python3

PROJECT_01 / 06 · FLAGSHIP

eligible

Patient-to-trial matching on harmonized clinical data. Harmonizes multi-source EHR into the OMOP CDM (dbt), converts free-text trial eligibility into structured logic, and runs audited agentic matching (LangGraph + DeepSeek) with per-criterion evidence. PHI redaction, RBAC, a tamper-evident audit chain, FHIR output, and a Next.js review UI.

610

Structured Criteria

Tests Passing

SHA-256

Audit Chain

MIT

Open Source

PythonFastAPILangGraphPineconedbtOMOP/FHIRNext.js

GitHub

# eligible patient-to-trial matching $ make all → Harmonize Synthea + legacy CSV into OMOP CDM (dbt) → 30 trials · 610 structured eligibility criteria → Agentic match: phenotype → retrieve → adjudicate → Verdict per criterion: eligible | ineligible | needs_review $ pytest -q ✓ 25 passed · audit chain · PHI redaction · matching

~/finreg-ml · python3

HuggingFace Demo

Open on HF ↗

PROJECT_02 / 06 · PUBLISHED

finreg-ml

Regulation-aware ML pipeline for finance. GovernedModel, SHAP explainability, fairness audits, EU AI Act compliance, drift detection (KS+PSI), consolidated reports. Published on PyPI.

Tests Passing

v0.2.0

PyPI Published

Modules

MIT

Open Source

Pythonscikit-learnSHAPPyPIGitHub ActionsFastAPI

GitHub PyPI ↗ Live Demo ↗

# finreg-ml v0.2.0 $ pip install finreg-ml Successfully installed finreg-ml-0.2.0 $ python -c "from finreg import GovernedModel; print(GovernedModel.__doc__)" EU AI Act compliant ML pipeline with SHAP explainability, fairness audits, drift detection (KS + PSI), and auto-reports. Supports: scikit-learn estimators | 10 modules | 46 tests $ pytest tests/ -ra ✓ 46 passed in 4.31s

~/clarify · typescript

AI Agent

Open in new tab ↗

PROJECT_03 / 06 · LIVE DEMO

clarify

Production LLM agent that turns a free-text brief into a validated BA artifact pack. Multi-step reasoning with a self-correcting clarification loop, typed IDs, RACI matrix, traceability. Llama 3.3 70B fallback chain via OpenRouter.

LLM AgentTool Calling

Next.jsAI SDKOpenRouterVercel

GitHub Live Demo ↗

# clarify $ # input: "vendor invoice approval tool" → Agent asks 6 clarifying questions before assuming → Restates assumptions, then ships artifact pack → Typed IDs (BR/SR/FR/NFR/TR/TC) · RTM · RACI register ✓ First-draft turnaround: ~6 hours → under 5 minutes

~/talent-rag · python3

PROJECT_04 / 06

talent-rag

Production-grade RAG system for talent intelligence. Hybrid retrieval (dense plus sparse), cross-encoder reranking, and ensemble generation over a FAISS index, served through FastAPI with a Streamlit console.

Hybrid

Retrieval

Cross-Enc

Reranking

FAISS

Vector Store

MIT

Open Source

PythonFastAPIFAISSCross-EncoderStreamlitRAG

GitHub

# talent-rag talent intelligence RAG $ python -m talent_rag.query "senior ML engineer" → Hybrid retrieval: dense (FAISS) + sparse (BM25) → Cross-encoder rerank → top-k candidates → Ensemble generation → grounded, cited answer $ uvicorn app:api · streamlit run ui.py ✓ FastAPI service + Streamlit console

~/agenteval · python3

CLI Framework

PROJECT_05 / 06

agenteval

AI agent evaluation framework with AgentRunner, LLM-as-judge scoring (OpenAI + Anthropic), CLI, safety checks for PII & prompt injection, multi-format export.

LLM-as-judgeSafety Checks

PythonOpenAIAnthropicpydanticCLI

GitHub

# agenteval v0.2.0 $ agenteval run --agent my_agent.py --suite evals/suite.json → Loading 11 modules · 62 tests passing → LLM Judge: claude-sonnet-4 (Anthropic) → Safety: PII scan ✓ · Injection scan ✓ → Export: JSON · CSV · Markdown ✓ Report saved: eval_results.json

~/crypto-stat-arb · python3

Backtest Engine

PROJECT_06 / 06

crypto-stat-arb

Statistical arbitrage engine. Engle-Granger cointegration, Johansen basket trading, Kalman filter hedge ratios, walk-forward backtesting, regime detection, paper trading via Kraken API.

Kraken APIKalman Filter

Pythonstatsmodelsscipypandas

GitHub

# crypto-stat-arb v0.1.0 $ python -m cryptoarb.backtest --pair BTC-ETH --window 90d → Engle-Granger cointegration test → Kalman filter hedge ratio (rolling) → Walk-forward backtest · regime detection → Market neutral confirmed (BTC corr ≈ 0.03) ✓ 107 tests · paper trading via Kraken API

04 · Tech Stack

The Toolbox04

// Foundation Models

Claude Opus 4.7Sonnet 4.6Haiku 4.5GPT-5GPT-4oo3Gemini 2.5DeepSeek R1Llama 4Grok 4

// Agents & Orchestration

LangGraphLangChainCrewAILlamaIndexMulti-Agent SystemsAgentic WorkflowsClaude CodeClaude Skills

// RAG & Retrieval

RAG PipelinesGraphRAGSemantic SearchText-to-SQLVector SearchDocument Intelligence

// Vector DBs & Graphs

PineconepgvectorNeo4jKnowledge Graphs

// AI Platforms & APIs

Anthropic APIMCPOpenAI RealtimeAzure OpenAIAWS BedrockGoogle Vertex AIHugging Face Hub

// Backend & APIs

PythonSQLFastAPIREST APIsMicroservicesSecure API Integration

// Eval, Safety & Governance

Guardrails AINeMo GuardrailsRagasTruLensLLM EvaluationRBACPHI MaskingHIPAA

// Cloud, DevOps & Data

AWSAzureGCPDockerKubernetesGitHub ActionsCI/CDPostgreSQLMongoDB

◉ Currently Learning

→ Reinforcement Learning → CUDA Programming → Rust → Transformer Fine-tuning → JAX/XLA → Options Pricing → RLHF → → Reinforcement Learning → CUDA Programming → Rust → Transformer Fine-tuning → JAX/XLA → Options Pricing → RLHF →

ATHARVA
JOSHI

About01

Projects That Ship02

Building in Public03

The Toolbox04

Education05

ATHARVAJOSHI

About01

Projects That Ship02

Building in Public03

The Toolbox04

Education05

ATHARVA
JOSHI