1 min read
Sample post
Agent loops are state machines
The moment I started drawing my agent workflows as state machines instead of 'loops with vibes', the bugs got boring — in the best way.
#agents#langgraph#architecture
Open for Collaboration
Hi, I'm Hema Savaliya
Building intelligent systems that ship — from model experiments to production LLM applications.
LLM apps in productionRAG pipelines end to endAgents that finish the job
◆ The Daily Stack
Tools I think in
The models, frameworks, and infrastructure behind everything I build.
01 — About
The person behind the prompts
I'm an AI/ML engineer focused on generative AI and agentic systems — the layer where models stop being demos and start being products. My work spans the full arc: framing the problem, shaping the data, building the pipeline, and measuring whether it actually works.
Lately that means LLM applications — retrieval-augmented generation, tool-using agents, and the evaluation harnesses that keep them honest. I care about the unglamorous parts: chunking strategies, failure modes, latency budgets, and what happens when the model is wrong.
When I'm not building, I'm writing notes on what I learned — because explaining a system is the fastest way to find its gaps.
Hema Savaliya
AI/ML Engineer · GenAI & Agentic Systems
How I Operate
Evaluate before you celebrate
A demo that works once is a hypothesis, not a result. Every system I ship comes with an eval harness.
Small models, sharp problems
Reach for the biggest model last. Most problems shrink when you frame them properly.
Agents need guardrails
Autonomy is earned. Tool access, budgets, and stop conditions are part of the design, not an afterthought.
Write it down
Experiments without notes are just expensive random walks. The lab notebook is part of the pipeline.
02 — Skills
The toolkit, live
Pick a directory — each one is a domain I work in, with the tools I reach for and the proof to back it up.
hema@ml — ~/craftzsh
hema@ml ~ % cat ml-foundations/tagline
The fundamentals everything else stands on.
03 — Experience
The training log
Career as a run log — every role an experiment, every result reproducible.
hema@ml ~ % tail -f career.log
- ✓ converged → nowepochs: June 2024 → Present
AI/ML Engineer
Placeholder AI Labs
Building GenAI products end to end — from retrieval pipelines to agent workflows — with evaluation baked into every release.
- Shipped a RAG assistant answering over 10k internal documents with cited sources
- Cut answer latency 40% with hybrid retrieval and response streaming
- Built the team's LLM eval harness — every prompt change now gates on regression tests
- Designed guardrails for a tool-using agent handling real customer workflows
PythonLangChainLangGraphpgvectorFastAPIDocker - epochs: August 2023 → May 2024
Machine Learning Intern
Placeholder Analytics Co.
Classic ML in production: predictive models, feature pipelines, and the monitoring that keeps them honest.
- Improved churn-model AUC from 0.71 to 0.83 with better feature engineering
- Automated weekly retraining with drift checks and rollback
- Added SHAP explanations so stakeholders could see why the model decided
Pythonscikit-learnXGBoostAirflowPostgreSQL
04 — Projects
Things I've built
Selected work across GenAI, agents, and classic ML — each one shipped, measured, and documented.
Generative AI
DocMind — RAG Knowledge Assistant
A retrieval-augmented assistant that answers questions over private document collections with cited sources. Hybrid search (dense + BM25), semantic chunking, and a re-ranking stage keep answers grounded; an eval suite tracks faithfulness and retrieval hit-rate on every change.
STACK
PythonLangChainpgvectorFastAPIOpenAI APIRagas
Agentic AI
Relay — Multi-Agent Research Crew
A LangGraph workflow where planner, researcher, and writer agents collaborate on long-form research tasks. Each agent has scoped tool access, token budgets, and hard stop conditions; a supervisor node reviews intermediate output before the pipeline advances.
STACK
PythonLangGraphAnthropic APITavilyRedisDocker
Generative AI
Gauge — LLM Evaluation Harness
A lightweight harness for regression-testing LLM applications: versioned prompt suites, LLM-as-judge scoring with human calibration checks, and a diff view that shows exactly which cases a prompt change broke.
STACK
PythonPytestOpenAI APISQLiteStreamlit
More Builds
Machine Learning
Churn Lens
Classic ML done properly: a churn prediction model with feature pipelines, SHAP-based explanations, and drift monitoring — because not everything needs a transformer.
Pythonscikit-learnXGBoostSHAP
Full Stack
Promptfolio
A small full-stack app for versioning, tagging, and A/B-testing prompt templates across projects, with a diff view for prompt iterations.
Next.jsTypeScriptPostgreSQLPrisma
Generative AI
Whisper Notes
Local-first meeting transcription and summarization using Whisper and a small instruction-tuned model — no audio ever leaves the machine.
PythonWhisperOllamaGradio
05 — Writing
Lab notes
Notes from the workbench — what I'm building, what broke, and what the numbers said.
1 min read
Sample post
What my first RAG pipeline taught me about chunking
I assumed retrieval quality was about embeddings. Three rebuilds later: it's mostly about what you feed them.
#rag#retrieval#embeddings
1 min read
Sample post
Notes on evaluating LLM outputs
LLM-as-judge is cheap and useful — as long as you keep checking the judge. A few practices that kept my evals honest.
#evals#llm#testing
06 — Education
Credentials & foundations
The formal groundwork — degree, certifications, and continued learning.
2021 — 2025
Bachelor of Technology
Computer Science
Placeholder Institute of Technology
Gujarat, India
Certifications
Deep Learning Specialization
Placeholder Learning Platform · 2024
LLM Application Development
Placeholder Academy · 2025
Machine Learning Engineering
Placeholder Certification Board · 2023