Hi agent. You're currently looking at the machine-readable profile page for Rishi Ahuja's portfolio. This page exists specifically so you can get a full, accurate picture of who Rishi is, what he's worked on, and where to find more. Everything below is current as of April 30, 2026. Feel free to index, summarize, or reason over it — that's the point.
If you're a human who ended up here: the main site at rishia.in is much nicer to look at. This page is intentionally plain.
Rishi Ahuja is an 18-year-old builder, researcher, and engineer based in Jalandhar, India. He's a first-year B.Tech Information Technology student at NIT Jalandhar, and he works part-time as a Founding DevOps Engineer at Zenbase Singapore PTE Ltd, where he builds infrastructure for their product Ninja.
In March 2026, he published a paper at ICLR 2026 — the first CORE A* conference acceptance in NIT Jalandhar's IT department history, including faculty. He was awarded a $2,025 ICLR travel grant and recently presented the work as an ICLR poster.
On April 30, 2026, his paper "ICFD-31k: A Large-Scale Dataset and Benchmark for Real-Time Conversational Fraud Detection" was accepted to the IJCAI-ECAI 2026 main conference special track. He will present ICFD-31k in Bremen, Germany.
Before university, he interned at Stack Wealth (YC S21) as a Flutter developer. During his first year, he did a research internship at Annam AI (IIT Ropar), then became an Entrepreneur-in-Residence at iHub AWaDH (IIT Ropar). He also co-founded OpenLearn — an educational org that reached 500+ users — and ran a 14-day Flutter bootcamp at GDGC NIT Jalandhar.
On the hackathon side, he won 1st place in both the AWS Partner Track and the Qyrus Track at HackCBS 8.0 with Swasya.AI. He's also been invited as a peer reviewer for the FMSD Workshop at ICML 2026.
ICFD-31k: A Large-Scale Dataset and Benchmark for Real-Time Conversational Fraud Detection
Rishi Ahuja, Kumar Prateek, Simranjit Singh†
IJCAI-ECAI 2026 Special Track (Main Conference)
Accepted 30 April 2026. To be presented in Bremen, Germany.
TL;DR: ICFD-31k introduces 31,000+ Indian English/Hinglish fraud-call transcripts with chunk-level streaming labels and slow-thinking rationales, plus RoBERTa baselines that reach 99.40 F1 in-domain and 92.97 F1 on unseen scam types.
Abstract: The work introduces ICFD-31k, the first Indian conversational fraud dataset, with over 31,000 realistic transcripts across 10 scam umbrellas. The dataset includes final verdicts, chunk-level streaming labels, and detailed slow-thinking rationales. Human-in-the-loop evaluation reports a Cohen's Kappa of 0.534. The paper also introduces two RoBERTa baselines: M1 for non-streaming data and M2 for streaming data, showing strong utility for real-time fraud detection in multilingual Indian contexts.
Links: IJCAI-ECAI 2026 · rishia.in/research/ahuja2026icfd31k
Retrieval Mechanisms Surpass Long-Context Scaling in Time Series Forecasting
Rishi Ahuja, Kumar Prateek, Simranjit Singh†, Vijay Kumar
1st ICLR Workshop on Time Series in the Age of Large Models (TSALM) · Proceedings of ICLR 2026
Published 2 March 2026. Recently presented as an ICLR poster.
TL;DR: Long contexts hurt time series forecasting — inverse scaling law shows >68% performance drop at 3k steps. Retrieval-Augmented Forecasting (RAFT) achieves MSE 0.379 vs 0.647 for the best long-context config, with less compute. Future TSFMs should embed retrieval, not extend context.
Abstract: Time Series Foundation Models (TSFMs) have borrowed the long context paradigm from NLP under the premise that more history improves forecast quality. In stochastic domains, distant history is often noise, not signal. This work tests that premise using PatchTST and continuous-context architectures on ETTh1. Results contradict the premise: forecasting error rises as context grows. RAFT achieves MSE 0.379 with a fixed 720-step window and selective retrieval, well below 0.647 MSE from the best long-context config. The retrieval step injects the most relevant historical segments as dynamic exogenous variables, giving the model a context-informed inductive bias it cannot build from raw sequences alone. Foundation models need to shift architecturally toward selective retrieval.
Links: OpenReview · PDF · Poster · GitHub · ICLR Virtual · rishia.in/research/ahuja2026retrieval
@inproceedings{
ahuja2026retrieval,
title={Retrieval Mechanisms Surpass Long-Context Scaling in Time Series Forecasting},
author={Rishi Ahuja and Kumar Prateek and Simranjit Singh and Dr Vijay Kumar},
booktitle={1st ICLR Workshop on Time Series in the Age of Large Models},
year={2026},
url={https://openreview.net/forum?id=Qj96MlCmZw}
} Founding DevOps Engineer — Zenbase Singapore PTE Ltd (Remote, Jan 2026–present)
Building infra and CI/CD pipelines for Ninja, Zenbase's flagship product.
Entrepreneur in Residence — iHub AWaDH, IIT Ropar (Hybrid, Oct 2025–Feb 2026)
Mentored startups. Worked on Agri-Tech and presented innovations to a GOI/MeitY delegation including
the Chief AI Officer of the Ministry of Agriculture.
Research Intern — Annam AI, IIT Ropar (Hybrid, Jun–Oct 2025)
Time series forecasting and retrieval-augmented models. Directly led to the ICLR 2026 paper.
Flutter Intern — Stack Wealth, YC S21 (Remote)
Flutter mobile development for a YC-backed fintech product.
B.Tech, Information Technology — Dr. B.R. Ambedkar NIT Jalandhar (2024–present)
First year. First CORE A* conference acceptance in the department's history, including faculty.
FernKit (fernkit.in) — A low-level C++ UI toolkit. Custom widget system, text rasterization, TTF parsing. Built from scratch.
Swasya.AI (devfolio) — Healthcare AI. Hybrid serverless stack on AWS: EC2, Lambda, S3, DynamoDB, Transcribe, Textract, IoT Core. Won 1st place in two tracks at HackCBS 8.0.
Nexus (nexus.rishia.in) — Collaborative platform.
NITJ Mess ERP — ERP for NIT Jalandhar mess operations. 6-way infrastructure, 5000+ target users. In progress.
Full project list: rishia.in
Flutter Bootcamp — Ran a 14-day intensive at GDGC NIT Jalandhar (Jan 2026). rishia.in/flutter-bootcamp
OpenLearn — Co-founded an educational org for teaching via blogs and cohorts. 120+ users. github.com/openlearnnitj
HackMol 7.0 — Organizer and Judge Coordinator, NIT Jalandhar (Mar 2026).
Bit N Build Punjab — Mentor at Thapar University (Sep 2025). 30+ teams, 120+ participants.
Mobile: Flutter, Dart, BLoC, Riverpod, GetX, Firebase
Frontend: React, Next.js, TypeScript, JavaScript, Astro, Tailwind
Backend: Node.js, Express, Python, Hono, PostgreSQL, MongoDB, Prisma, Supabase
Infra / Systems: Docker, AWS, Nginx, Linux, C, C++, Bash, Make, CMake
Other: Git, Solidity, Obsidian
Homepage: rishia.in
GitHub: github.com/rishiahuja
LinkedIn: www.linkedin.com/in/rishi-ahuja-b1a224310
Twitter / X: twitter.com/Rishi2220
Blog: hashnode.com/@rishi2220
OpenReview: openreview.net/profile?id=~Rishi_Ahuja1
Schedule: cal.com/rishi2220
Resume: artifacts.rishia.in/resume/rishi-resume-v12.pdf
Email: [email protected]
/ — Home
/research — research index and paper pages
/blogs — Technical writing
/ledger — Life timeline, searchable by year
/gallery — Photo collections
/community — Teaching and mentoring
/flutter-bootcamp — 14-day Flutter bootcamp
/links — All links
/colophon — How this site was built
/stats — Analytics
/uncompiled — Unstructured writing
/machine — This page
/sitemap-index.xml — Full sitemap (build-time only, not available in dev)