Skip to content
Yash.Patil
Open to AI/ML engineering roles

Yash Patil

AI/ML Engineer ·

I build production AI systems — retrieval, agents, and evals — that hold up under real traffic.

Chat with YashView work
Portrait of Yash Patil
Telemetry Live
2M+
queries / day served
+15%
NDCG@10 ranking lift
<50ms
p95 ranking latency
~4 yrs
applied AI/ML

About

Applied AI engineer, eval-first.

I'm an AI/ML engineer with ~4 years of experience shipping LLM, retrieval, and agentic systems to production. Right now I'm an AI/ML engineer at Tip Top Technologies, where I redesigned a real-time ranking pipeline (NDCG@10 +15%, p95 120ms → <50ms) and built an embedding-based semantic retrieval system serving 2M+ queries a day.

My work runs the full stack of applied AI: RAG and hybrid search, multi-agent orchestration, and evaluation harnesses that actually catch hallucinations — much of it in healthcare, where correctness and HIPAA compliance are non-negotiable. I care about reproducibility and open weights: my projects are open-source and reproducible.

I hold an M.S. in Computer Science from Santa Clara University. Before that I built large-scale data and ML pipelines across search and life-sciences companies.

Education

  • M.S. Computer Science & Engineering

    Santa Clara University

    Sep 2023 – Jun 2025 · Santa Clara, CA

    GPA 3.63 · Artificial Intelligence, Algorithms, Operating Systems, Computer Architecture

  • B.Tech Computer Science & Engineering

    Rajarambapu Institute of Technology

    Aug 2016 – Oct 2020 · Sangli, India

    Data Structures, DBMS, Principles of Machine Learning

Experience

Where I've shipped.

  1. AI/ML Engineer · Tip Top Technologies

    Oct 2024 – Present · Sunnyvale, CA

    • Redesigned the real-time recommendation ranking pipeline (retrieval + re-ranking) for News360, lifting NDCG@10 by 15% and cutting p95 latency from 120ms to under 50ms via profiling, caching, and serving optimizations.
    • Built an embedding-based semantic retrieval system (FAISS + Pinecone) serving 2M+ queries/day with production monitoring, raising recall@100 by 12%.
    • Shipped low-latency AI microservices on Kubernetes/AWS with CI/CD, observability, and reliability hooks.
    • Building an agentic meeting copilot for COM360 — a voice interface backed by graph-based long-term memory, gated by explicit consent and guardrails.
    PythonFAISSPineconeKubernetesAWSLLMsCI/CD

  2. AI Intern · Samvid

    Jul 2024 – Sep 2024 · Remote

    • Built LLM evaluation pipelines benchmarking quality and latency across GPT-4o, Gemini, and AWS Bedrock models (Claude, Llama 3.1, Mistral) to inform production-readiness decisions.
    • Automated data-validation workflows (schema, nulls, distributions) for ML pipelines, accelerating iteration ~40% and cutting cross-team syncs by ~6 hours/week.
    • Added real-time monitoring and structured logging for model evaluation and deployment stability.
    GPT-4oClaudeAWS BedrockLangChainRAGPython

  3. Data Scientist · Searchspring

    Apr 2023 – Aug 2023 · Remote

    • Developed spell-correction NLP models that improved search relevance by 30%.
    • Built large-scale Databricks PySpark ETL pipelines, boosting analytics throughput by 40%.
    PySparkDatabricksNLPPython

  4. Data Scientist · Anju Life Sciences Software

    Mar 2021 – Dec 2022 · Pune, India

    • Led HIPAA-compliant analytics across clinical-trial, claims, and EHR datasets; raised ingestion throughput by 45% with optimized ETL pipelines and SQL Server stored procedures.
    • Built REST APIs to deliver ML-assisted insights.
    SQL ServerREST APIsHealthcare

Projects

Things I've built.

Open-source and mostly AI-native. The featured projects power the live demos arriving on this site.

MCP Repo Agent
Featured
A live, in-page agent that explores any public GitHub repository over a real in-process MCP server — reading the tree, README, and files through scoped read-only tools and streaming each tool call as it works. Built for this site.

Live MCP agent

TypeScriptMCPLLMAgents
CodeLive demo
Open-Scribe
Featured
Open-source, FHIR-native ambient medical scribe: encounter audio → SOAP note → ICD-10/CPT → FHIR bundle. Ships a real evaluation harness on the PriMock57 dataset.

FHIR + real eval harness

PythonWhisperFHIREvalsLLM
Code
ClinEval
Featured
Measures clinical hallucination, failure modes, and demographic equity in LLM-generated medical notes using an LLM-as-judge pipeline over synthetic data.

Hallucination scoring

PythonLLM-as-JudgeEvalsHealthcare AI
CodeDemo
Voice-Triage
Featured
Voice-first clinical intake agent with red-flag triage — Whisper STT + Llama 3.3 (Groq) + browser TTS, deployed on Vercel. An open take on the real-time voice agent stack.

Whisper + Llama 3.3, on Vercel

Next.jsGroqLlama 3.3WhisperVoice AI
CodeDemo
Circuit-Extract
Featured
Extracts structured data from electrical circuit diagrams into machine-readable relational graphs — a hybrid vision pipeline of PaddleOCR + Gemini/Claude Vision + OpenCV.

Vision → structured graph

PythonVisionOpenCVMultimodal
Code
Databricks Smart Merge
Featured
Reproduces a real Databricks Assistant failure where partial LLM fixes overwrite working cells, then applies fixes as a safe patch with an inline diff preview — practical LLM safety in code workflows.

LLM code safety

PythonStreamlitGemini
Code
AutoLabel
Featured
Autonomous, self-improving data-labeling system: LLM-powered weak supervision with a Karpathy-style ratchet that ratchets quality up over iterations.

Self-improving weak supervision

PythonWeak SupervisionLLM
Code
Fintech Decisioning Agent
A LangChain agent on AWS Bedrock with RAG workflows and a FastAPI backend, with an EKS deployment pipeline (CI/CD via GitHub Actions).
LangChainAWS BedrockFastAPIKubernetes
Code

Ask my AI

Chat with Yash

A retrieval-grounded agent that answers questions about my work — running on open-weight models behind a secured proxy.

CHAT-RAG Ready

Open-weight RAG · grounded in this site's content

Open-weight model + pgvector RAG through a rate-limited, bot-checked proxy. No API keys in the browser.

MCP agent

Explore any repo

Paste a public GitHub repo and a tool-using agent explores it over MCP — listing the tree, reading the README and key files — then explains what it does and how it's built. Its tool calls stream live.

MCP-AGENT Ready

Tool-using agent · real MCP tools over the GitHub API · public repos only

Read-only, single-repo tools through a rate-limited, bot-checked proxy. Repo contents are treated as untrusted data. No API keys in the browser.

Clinical eval

Score a note for hallucinations

Paste a synthetic clinical note. An extraction pass produces a structured, FHIR-ish summary (problems, medications, plan); a separate grounding pass labels every extracted claim grounded, partial, or unsupported and cites a span. Then code — not the model — verifies each span against the note and computes the score, so the number can't be gamed. The methodology is shown, not hidden.

CLINEVAL Ready

Use synthetic data only — never real patient information (PHI). Your note is processed in memory and is not stored or logged by this site; it is sent to a model provider for processing only.

Two-pass eval · structured extraction + per-claim grounding · score computed in code

Runs through the same rate-limited, bot-checked proxy as the other demos. The note is treated as untrusted data, kept in memory only, and never logged or persisted. No API keys in the browser.

Agent Skills

Reusable Agent Skills

Two production-grade Anthropic Agent Skills (SKILL.md) that package the eval methods behind the demos above — runnable, spec-faithful, and MIT/Apache-licensed. Each is a faithful, standard-library-only Python port of real, shipped work.

clinical-note-eval
MIT
Two-pass extract-then-ground eval for clinical-note extractions, with a deterministic, code-side span check the model can't inflate.

Packages: The methodology behind the live “Score a note” demo (ClinEval + Open-Scribe).

PythonEvalsGroundingHallucinationLLM-as-judge
View skillLive demo: Score a note
token-eval-harness
Apache-2.0
Token-level multiset-F1 scorer for free-text output vs references — the honest surface-overlap metric, per-example and macro-aggregated.

Packages: Open-Scribe's PriMock57 / ACI-Bench token-F1 evaluation harness.

PythonToken-F1EvalsRAGPriMock57
View skillOpen-Scribe

Anthropic SKILL.md format — clone one into ~/.claude/skills/. Each runs its worked example and tests offline, with no API key. Browse the repo

Skills

Tools I reach for.

Languages

  • Python
  • TypeScript
  • JavaScript
  • SQL
  • Java
  • Bash

LLMs & Agents

  • LangChain
  • LlamaIndex
  • RAG
  • Multi-Agent
  • MCP
  • Prompt Engineering
  • Evals

Models

  • GPT-4o
  • Claude
  • Gemini
  • Llama
  • Mistral
  • gpt-oss

Retrieval & Vector DBs

  • FAISS
  • Pinecone
  • pgvector
  • Chroma
  • BGE / nomic embeddings

ML & Data Science

  • PyTorch
  • TensorFlow
  • scikit-learn
  • Transformers
  • NLP
  • Forecasting

Data Engineering

  • Spark / PySpark
  • Databricks
  • Kafka
  • BigQuery
  • ETL/ELT

Cloud & MLOps

  • AWS (Bedrock)
  • GCP (BigQuery)
  • Azure OpenAI
  • Docker
  • Kubernetes
  • CI/CD

Visualization

  • Streamlit
  • Tableau
  • Power BI
  • Matplotlib
  • Plotly

Contact

Let's build something.

Open to AI/ML engineering roles and genuinely hard problems. The fastest way to reach me is email.

Email

Email me