AI Engineer · Agent Systems · LLM Infrastructure
I build AI systems that ship — not demos. EquityLens (healthcare bias detection) is in production with 92.9% R² and 80%+ accuracy. Currently building ContextIQ, a drop-in LLM context/cost optimizer, and Neuron OS, a local-first OS for autonomous agents.
Author of sentinel-cli — local-first AI code reviewer on npm.
Ahmedabad, India · AI Automation Intern @ Phaze AI · 4th Year CS @ Indus University (CGPA 9.00)
I build tools at the intersection of LLMs and production systems — agent infrastructure, evaluation frameworks, and retrieval, not chatbot wrappers.
At Phaze AI, I work on multi-agent automation workflows in production. EquityLens, a healthcare bias detection system, is live with measured results (92.9% R², 80%+ accuracy) — not a notebook demo.
Right now I'm deep in:
- LLM context/cost optimization (ContextIQ) — reducing token spend at the proxy layer
- Multi-agent orchestration and agent-to-agent IPC
- MCP (Model Context Protocol) and agent tooling
- LLM evaluation — because "it feels right" isn't a metric
- Vector search internals (built 9 ANN algorithms from scratch to understand why)
My development philosophy: ship first, refine second, measure everything.
Languages
Frontend
Backend
AI/ML
Vector Databases
Pinecone ChromaDB FAISS Weaviate
Cloud & DevOps
Drop-in proxy that cuts LLM context/token spend without touching your app logic. Full spec, TDD test suite, and dashboard built out — GTM model priced on % of token savings delivered.
Why it matters: Token cost is the biggest hidden line item in production LLM apps. ContextIQ optimizes the context window before it hits the model.
Stack: TypeScript · Node.js · Proxy architecture · Multi-LLM APIs
Status: Product spec + TDD suite complete. Building toward pilot.
A system that detects bias in healthcare ML models, live and running — not a portfolio demo.
Results: 92.9% R² · 80%+ accuracy
Stack: Python · ML pipeline · FastAPI
A local-first, TypeScript-native operating system for autonomous AI agents — spawn typed agents, run them across terminal, web, chat, and API surfaces, with built-in audit logging, per-agent tool policies, cost attribution, and an encrypted credential vault.
Why it matters: Most agent frameworks are GUI-first or notebook-first. Neuron OS is built for composability — piping agents like Unix commands, with production controls (RBAC, sandboxing, distributed workers) baked in, not bolted on.
Stack: TypeScript · Bun · 13 AI providers · 11 platform adapters · MCP
Status: Active development. → Repository
Your codebase, reviewed locally. No SaaS. No data leaks. Your keys, your rules.
Problem: Existing code review tools are either cloud-dependent, expensive, or shallow.
Solution: A local-first CLI with 13+ specialized analyzers covering security, accessibility, TypeScript strictness, React patterns, and dependency vulnerabilities. Plugs into any LLM (OpenAI, Claude, Groq, Gemini).
Stack: Node.js · TypeScript · Multi-LLM APIs · Docker
Key Features:
- SQL injection, XSS, secrets detection
- WCAG 2.1 accessibility checks
- GitHub PR integration via CLI
- SARIF output for GitHub Security tab
- Auto-fix for common issues
- Works 100% offline for static analysis
Nine ANN algorithms. Hand-rolled. No shortcuts.
Problem: Most developers use vector DBs as black boxes. I wanted to understand what's actually happening inside.
Solution: A production-grade vector database built from scratch in Python featuring nine hand-rolled ANN algorithms (HNSW, IVF, PQ, LSH, and more), a full RAG pipeline, multimodal ingestion, FastAPI server, and Kubernetes deployment configs.
Stack: Python · FastAPI · NumPy · Docker · Kubernetes
Key Features:
- 9 ANN algorithms implemented from scratch
- RAG pipeline with multimodal support
- REST API via FastAPI
- K8s deployment configs for production
- Benchmarking suite across algorithms
Ancient script. Modern tokenization. BPE from scratch.
Problem: Sanskrit is morphologically complex. No production-grade tokenizer existed for GPT fine-tuning workflows.
Solution: Custom BPE tokenizer built from scratch. Fine-tuning pipeline for GPT-3.5/4 on Devanagari text. Multilingual support across Sanskrit, Hindi, English.
Stack: Python · PyTorch · Transformers · OpenAI API
Stop applying blindly. Know your fit score before you apply.
Problem: Job seekers waste hours applying to roles they're underqualified or overqualified for.
Solution: An intelligent agent that parses resumes with NER, semantically matches them to job descriptions, surfaces skills gaps, and gives ATS optimization suggestions.
Stack: Python · spaCy · Transformers · FastAPI · React · PostgreSQL
What's standing between you and your target role?
Problem: Developers don't know which skills to prioritize for career growth.
Solution: TypeScript-based skills gap analyzer that maps your current skills against target job requirements and surfaces the highest-ROI things to learn next.
Stack: TypeScript · Node.js · LLM APIs
Because good prompts deserve version control.
Stack: TypeScript · React · Node.js
| Topic | Status |
|---|---|
| LLM context/cost optimization | Active |
| MCP (Model Context Protocol) | Active |
| Agent memory architectures | Active |
| LLM evaluation frameworks | Active |
| Multi-agent coordination patterns | Active |
| RAG pipeline optimization | Ongoing |
| Vector DB internals | Shipped ✅ |
| Distributed systems fundamentals | Ongoing |
I push when things work, not when they're perfect. Readmes matter. Tests matter more. If a project has no README, it might as well not exist.
I prefer small, focused repos over monorepo sprawl. I'd rather have five sharp tools than one giant project that tries to do everything.
Commits tell a story — I try to make them readable.
| 🌐 Portfolio | kunjshah.vercel.app |
| linkedin.com/in/kunjshah05 | |
| kunjkshahdeveloper@gmail.com | |
| 🐦 X / Twitter | @kunjshah_dev |
Building in public. If you're working on agent infrastructure, evals, or retrieval — reach out.