Aryan Sahni · CS, UC Santa Cruz · Cum Laude
BuildingAIsystemsthatthink,speak,andremember.
Engineering real-time voice AI and agentic systems. Founding Engineer at Aura, a South Park Commons affiliate. Previously Snowflake.
Selected projects · 2024 — 2026
Things I've built.
Voice-first AI companion app with cycle-phase awareness and persistent semantic memory. Architected the full real-time pipeline: Groq Whisper STT → LLaMA 3.3 70B → ElevenLabs Flash v2.5 TTS → LiveKit WebRTC, with phase-conditioned memory retrieval using pgvector composite scoring. Built as a South Park Commons affiliate, SF.
- React Native
- Supabase
- pgvector
- Groq
- ElevenLabs
- LiveKit
- Expo
Trust layer for AI coding agents — verifies completion claims against real process exit codes instead of self-reports. Blocks writes to protected paths before they hit disk and records every event in a SHA-256 hash-chained ledger, surfaced as a local trust-score dashboard and Claude Code plugin.
- TypeScript
- Node.js
- MCP
- Claude Code
- pnpm
- Vitest
8-agent ReAct system collapsing 30-minute outreach into 60 seconds. Resume gap analyzer, warm-path finder, and 0–100 response-rate scorer.
- React
- FastAPI
- Groq
- Apollo.io
- Gmail API
Autonomous AI software factory — ideation through deployment on ECS Fargate. Six self-healing maintenance agents replace PagerDuty with semver-aware auto-merge.
- Next.js
- AWS ECS
- Nemotron 120B
- PostgreSQL
Full-stack agent reviewing GitHub PRs in under 60s with line-level inline comments. Resilient inference layer with model failover and self-healing JSON parser.
- Next.js
- FastAPI
- GitHub API
- NVIDIA NIM
Multi-agent stock analysis tool. Enter a ticker and 8 specialized agents — price, news, sentiment, technicals, fundamentals, bull case, bear case — stream live over SSE before a Portfolio Manager agent delivers a BUY/HOLD/SELL verdict with confidence score.
- Python
- Flask
- React
- Claude
- Finnhub
- SSE
- SQLite
ML platform that predicts football player fair market value and surfaces transfer mispricings. Trained a stacked ensemble (XGBoost + LightGBM + RF) on 2.7M+ rows of 30-year transfer data — R²=0.37, MAE €3.1M; similarity search across 17,596 player profiles runs in under 100ms.
- XGBoost
- LightGBM
- FAISS
- SHAP
- FastAPI
- TanStack Start
- Python
Experience
Where I've worked.
AI Research Scientist · Jul 2026 — Present
University of Washington, Department of Radiology
- — Building AI systems that draft chest X-ray reports with vision-language models in a human-in-the-loop clinical workflow.
- — Evaluating and fine-tuning open-source multimodal LLMs (MedGemma, CheXagent) with LoRA/QLoRA, and architecting a multi-stage inference pipeline for structured findings detection and report generation to reduce hallucination.
- — Designing the end-to-end AI system — problem scoping, model selection, and data curation pipelines — optimizing for dual-GPU compute constraints, generalization across data distributions, and evaluation rigor.
Currently
Shipping Aura. Somewhere with good coffee.
Avg voice response latency
~800ms
Current LLM
Llama 3.3 70B
Embedding model
text-embedding-3-small
Retrieval architecture
Hybrid RAG · pgvector
Toolkit
What I work with.
Languages
- Python
- TypeScript
- JavaScript
- SQL
- Bash
AI & ML
- LLM integration
- RAG
- pgvector
- Multi-agent orchestration
- Agentic system design
- scikit-learn
- Pandas
- NumPy
- Predictive modeling
- Statistical modeling
- Feature engineering
- Model evaluation
Data & Infrastructure
- Snowflake
- Snowpark
- Docker
- Vercel
- Render
- AWS ECS Fargate
- GitHub Actions
Frontend & Mobile
- React
- React Native
- Expo
- Next.js
- Tailwind CSS
- Zustand
- TanStack Query
Backend & APIs
- FastAPI
- Flask
- REST APIs
- Server-Sent Events
- GitHub API
- GraphQL
- LiveKit
- WebRTC
- PostgreSQL
- Supabase
Daily Tools
- Claude Code
- Cursor
- GitHub Copilot
- ElevenLabs
Contact
Let's build something worth remembering.
Open to SWE and AI/ML roles. Building agents, voice systems, and the infrastructure between them.