SAI ARYAN GOSWAMI

Production AI systems, mostly full-stack. The current work is mutual fund analytics for a wealth management client, an audio intelligence platform that's live with paying infrastructure behind it, and a voice agent that makes real phone calls. The pattern is usually the same: a workflow someone is doing badly by hand, and a system that takes it over end to end.

Next.js, React and TypeScript on the front, FastAPI and Python on the back, Postgres with Neon and pgvector for data. For AI work: Groq, Claude, OpenAI, AssemblyAI, plus Langfuse for observability and LangGraph when a pipeline needs structure. He deploys on Vercel and Render, and works on Arch Linux in Neovim.

Yes, for select freelance work. He's running several client engagements right now across fintech, SaaS, insurance and voice AI. Rates and scope are worth an email rather than a chat answer: saiaryan.goswami1784@gmail.com.

It placed 2nd globally in the World Bank track of Hack-Nation's 5th Global AI Hackathon, out of around 300 teams, built in 18 hours with Amit Kumar Gupta. It maps skills for people in the informal economy, where real ability is invisible to formal systems. Three engines with different cost profiles: local ONNX embeddings with pgvector, a NetworkX graph seeded from ESCO that ranks adjacent skills by wage premium and automation resilience, and an LLM last so most work never reaches it.

Mutual fund portfolio analytics for a wealth management client, live at folio-mvp-ten.vercel.app. It ingests the Consolidated Account Statement PDF, calculates XIRR with a hand-written Newton-Raphson solver that falls back to bisection when it diverges, and flags the thing that quietly costs investors money: holding a regular plan when the direct plan of the same fund exists. The upload screen is live; the results screen is still ahead of the API.

An audio intelligence platform, live at audioanalyzer.saiaryan.in with 15 users. AssemblyAI transcribes, Groq LLaMA-3 analyses, and the interesting part is that the evaluation rubric is a plugin rather than an assumption. Rosenshine for teaching, SPIN for sales calls, STAR for interviews, and adding a fourth doesn't touch core logic.

Yes. UNMAPPED runs fastembed with ONNX for local CPU embeddings and pgvector on Neon for retrieval. He also built a RAG chatbot SDK during his internship at Sarvam AI. He tends to argue against reaching for a vector database when the corpus is small enough to fit in context, which is more often than people expect.

He was a Full-Stack Developer at Yutrp in Gurugram, building Next.js applications serving 1,000+ users and an OCR pipeline at 93% accuracy, then Frontend Developer on the API Engineering team at Sarvam AI in Bengaluru, working on workflow automation and a RAG chatbot SDK. He went freelance after that.

Noida, in Delhi NCR, India. He works with clients remotely, including in the UK and the US.

B.Tech in Computer Science with a Data Science specialisation from ABES Engineering College in Noida. He also co-authored a paper on a blockchain e-voting system, presented at ICETCI 2026.

Yes, and it caught a real bug. Langfuse traces on AudioAnalyzer surfaced that the analysis was sometimes scoring transcript segments without correctly attributing them to the right speaker, so feedback landed on the wrong person's words. Nothing errored and the output looked fine, which is exactly why per-run tracing was the only way to see it. He gave a lightning talk on this at GDG Noida.

Email is fastest: saiaryan.goswami1784@gmail.com. He's also on GitHub at SaiAryan1784, on LinkedIn, and on X as NUCLEAR1784.