Priyanshu Gupta
Engineering systems for decision-making. Product, AI, and finance.
About
I build systems people bet money on, which is a good way to learn the difference between code that works and code that is right. Exchange infrastructure that settles real funds, AI agents that have to be correct rather than merely convincing, and the unglamorous edge cases that decide whether either one survives a Tuesday. First tech hire at Krypto, founding engineer at TanX back when the tech team was two people and a lot of optimism, and now Prysm. I teach what I build: 10,000+ people on YouTube and an eight-week backend architecture course at GeeksForGeeks. There is always something half-finished on my machine. The parts worth explaining end up here.
Numbers
- People mentored
- 10K+
- GFG instructor rating
- 5/5
- On Topmate
- Top 1%
Work
Prysm Finance
Dec 2024 – PresentFounder & Chief Architect
An AI-first financial research and portfolio intelligence platform. In its first year: 30,000+ monthly visits, $22M of connected portfolios, 5,000+ users, and 50,000+ AI queries answered. Prysm covers the whole investor journey, from pre-trade research and screening, through in-trade monitoring, to post-trade portfolio analysis, in a single conversational interface. It automates the fundamental, technical, and qualitative work so conviction gets built early, when it still matters, and cuts research time by more than half. We built it from idea to production: natural language stock screening, AI-generated company research that shows its sources, real-time news with sentiment analysis, quantitative portfolio metrics, and a voice assistant that answers in 11 languages.
TanX (prev. Brine Finance)
Nov 2022 – Jan 2024Senior Product Engineer & Founding Team Member
Joined when the tech team was two people and helped grow it to ten, mentoring engineers and owning the exchange's core end to end: the order execution engine, on-chain deposits, L1 and L2 transfers, and StarkEx zero-knowledge settlement. TanX became a top-10 global DEX and cleared over $1B in quarterly trading volume. I led the integrations with institutional market makers, LayerSwap, and Merkle Science, which is where most of the liquidity and compliance surface came from. When a StarkEx settlement ratio error stalled more than 1,000 transactions, I spent six weeks reverse-engineering their fee and price model and wrote the automation that cleared every edge case, unblocking the trading loop.
GeeksForGeeks
Oct 2023 – Dec 2023Course Instructor (Part-Time)
Designed and taught an eight-week course on scalable backend architecture in Python, taking working engineers from single-server thinking to systems that survive load. Rated 5/5 across every delivery metric.
Krypto
Aug 2021 – Oct 2022Product Engineer, First Tech Hire
First engineer in. Owned the Portfolio feature from concept to production, designing for eventual consistency with event-based async computation and reconciliation that was idempotent and retried cleanly. Tripled database throughput with PgBouncer pooling in front of PostgreSQL, and added Datadog logging and APM so failures could be seen rather than guessed at.
Writing
- Integrating Google Sign-In with Django and ReactJS
- Everything about Celery
- Dockerizing Django with Traefik, Let's Encrypt, Celery, Celery Beat, Redis, and RabbitMQ
- Decoding Notifications Push, In-App, Email, SMS, and More
Projects
Portfolio X-Ray
2026You own four funds, so you are diversified. Except three of them hold the same five large-caps, and you are far more concentrated than your broker's neat little list suggests. That hidden overlap is exactly what hurts in a downturn. Portfolio X-Ray opens every Indian mutual fund and ETF, pulls out the stocks actually inside, and adds them up, so you can see where your money really sits by sector, market cap, and geography. Reads your holdings from Zerodha, broker-agnostic by design, and never places an order.
- Python
- Zerodha Kite API
- No build step
- MIT
LoRA and Efficient LLM Serving for Financial Expert Agents
2025How do you serve a hundred specialised financial models without buying a hundred GPUs? A research survey of parameter-efficient fine-tuning and multi-tenant serving: LoRA training 10,000x fewer parameters, and where vLLM, S-LoRA, and dLoRA actually earn their claimed 4x throughput rather than just quoting it. Evaluated FinLoRA, FinGPT, and layered RAG agent architectures for financial reasoning.
- LoRA
- vLLM
- S-LoRA
- RAG
AI Trust & Citizen Participation in Digital Public Services
2025Does an AI that sounds human earn more trust than one that works? We ran a 2x2 experiment on 143 people to find out. Both helped, but sounding authentic stopped counting for much once the thing performed badly. The odd result: trust predicted whether people would participate, and predicted nothing at all about whether they would hand over their data. People weigh how sensitive the information is over how much they like you.
- Qualtrics
- Prolific
- Statistics
LetsProgressify
2024A learning platform built for a 10,000+ subscriber community, supporting structured resource publishing and content monetisation. Included an ETL pipeline that aggregated hackathons, competitions, and conferences from several platforms into one tracked dashboard.
- Python
- Django
- Next.js
- PostgreSQL
- Docker
Privacy-Preserving Computation Techniques
2024Can several parties compute on data none of them are allowed to see? Implemented partially homomorphic encryption, secure multi-party computation, and zero-knowledge proofs across a multi-server Flask setup, then measured what each one actually costs. The cryptography is elegant; the bill is real.
- Python
- Flask
- Cryptography
- ZKP
Education
- 2024 – Present
Georg-August-Universität Göttingen
MSc Computer Science, Spec. Applied System Engineering with minor in Business Information Systems
- 2018 – 2022
Vellore Institute of Technology
B.Tech, Information Technology
Contact
If you're working on something interesting, or you want to argue with something I've written, I'd like to hear about it.