# Rishi Ahuja > Undergraduate researcher at NIT Jalandhar, building and evaluating reliable AI systems, and an incoming intern at Cisco. The complete structured profile, with abstracts, experience, teaching, and contact details, is at https://rishia.in/machine. ## Research - [It’s Not What the Scammer Says, It’s How You Reply: Small Scam Detectors Judge Code-Mixed Fraud Calls by the Victim](https://rishia.in/pub/ahuja2027scambranch): AAAI-27 Student Abstract. Across 2,643 code-mixed calls, a single victim reply splits the verdicts of 14 small open models in 30–80% of conversation sets. Fine-tuning Qwen2.5-1.5B raises macro-F1 from 0.73 to 0.86 and roughly halves these splits, but scams where the victim cooperates stay the hardest to catch. - [DoubtBench: Does Self-Doubt Help Language Models—or Just Make Them Check?](https://rishia.in/pub/ahuja2027doubtbench): Findings of WSDM 2027. Across 423,600 responses, adding self-doubt wording beyond checking changes accuracy by −1.43 to +2.59 percentage points. Its effect depends on the model, task, and instructions already present. - [Do LLM Reviewers Respect Scope? ScopeBench-PR: A Benchmark for Scope Fairness in Peer Review](https://rishia.in/pub/ahuja2026scopebenchpr): GlobalSouthAI Workshop · IJCAI–ECAI 2026. ScopeBench-PR audits LLM peer reviewers for scope fairness using counterfactual prestige and language-of-study variants. Regional penalties are model-dependent, scope-aware prompting helps unevenly, and rebuttal often leaves a sticky score penalty. - [ICFD-31k: A Large-Scale Dataset and Benchmark for Real-Time Conversational Fraud Detection](https://rishia.in/pub/ahuja2026icfd31k): IJCAI–ECAI 2026 · AI for Social Good Special Track. 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. - [Retrieval Mechanisms Surpass Long-Context Scaling in Time Series Forecasting](https://rishia.in/pub/ahuja2026retrieval): ICLR 2026 · TSALM Workshop. Long contexts hurt time series forecasting by adding noise (inverse scaling, >68% worse at 3k steps), while selective retrieval (RAFT) beats them with lower MSE (0.379 vs 0.647) and less compute - future TSFMs should embed retrieval instead. ## Writing - [Towards the Modern Transformer Architecture](https://rishia.in/writings/towards-the-modern-transformer-architecture): Tracing how the 2017 Transformer evolved into the recipes used by modern frontier models: norms, activations, attention variants, RoPE, KV cache, and more. - [You Don't Know WebSockets. Yet.](https://rishia.in/writings/you-dont-know-websockets-yet): The WebSocket protocol, handshake to frames, and how bidirectional sockets actually work. - [Go Beneath the Abstraction: Building Interactive UIs with FernKit](https://rishia.in/writings/go-beneath-the-abstraction-building-interactive-uis-with-fernkit): Building interactive UIs from scratch in C++ with FernKit’s renderer and widgets. - [Shamir's Secret Sharing Scheme and Multi Party Computation.](https://rishia.in/writings/shamirs-secret-sharing-scheme-and-multi-party-computation): Shamir’s secret sharing and multi-party computation for splitting a private key. - [Your Hardest "Hello World!": Text Rasterization #1](https://rishia.in/writings/your-hardest-hello-world-text-rasterization-1): How TrueType fonts are parsed and pixels get onto the screen. - [Bits of Trust: The Elegance of AES](https://rishia.in/writings/bits-of-trust-the-elegance-of-aes): How AES encrypts a block, from S-boxes through the last round. - [Building Rosenblatt's Perceptron From Scratch in Flutter](https://rishia.in/writings/building-rosenblatts-perceptron-from-scratch-a-comprehensive-technical-deep-dive): Building Rosenblatt’s perceptron in Flutter, with the geometry drawn out. - [Getting Cracked at Clean and BLoC Architecture](https://rishia.in/writings/getting-cracked-at-clean-and-bloc-architecture): Clean architecture and BLoC in Flutter, past the introductory examples. - [Getting Started at BLoC Architecture](https://rishia.in/writings/getting-started-at-bloc-architecture): The BLoC pattern in Flutter, with small examples you can run. - [Resource Management with Probabilistic Scheduling in Linux](https://rishia.in/writings/resource-management-with-probabilistic-scheduling-in-the-context-of-linux): How the Linux scheduler allocates CPU, and why probabilistic methods show up. - [State of the Art - ART (Android Runtime)](https://rishia.in/writings/art): How Android Runtime compiles, garbage-collects, and changes what apps can assume. - [Comprehensive Arch Linux Blog](https://rishia.in/writings/comprehensive-arch-linux-guide): Installing, configuring, and administering Arch Linux from a blank disk. ## Blurbs - [Health Update.](https://rishia.in/blurb/health-update): The one thing that’s done more for me than anything else. - [My IJCAI–ECAI 2026 and Germany Experience in Bremen](https://rishia.in/blurb/my-ijcai-ecai-2026-and-germany-experience-in-bremen): Two research presentations, a last-minute proxy talk, two awards, a lot of people, and a week of figuring out Bremen in between all of it. - [My interview experience with Cisco](https://rishia.in/blurb/cisco-interview): Campus hiring in September 2026, from the OA to the final HR round. - [My ICLR 2026 and Brazil Experience in Rio de Janeiro](https://rishia.in/blurb/iclr-2026-rio-de-janeiro): A long route to Rio, a paper presentation at ICLR, and the practical lessons that came with figuring out Brazil in real time. - [Building Fern](https://rishia.in/blurb/fern): The complete journey of developing a modern web graphics library ## Profiles - [Machine-readable profile](https://rishia.in/machine) - [Résumés](https://rishia.in/resume) - [Google Scholar](https://scholar.google.com/citations?hl=en&user=EZff3KsAAAAJ) - [ORCID](https://orcid.org/0009-0000-3403-010X) - [OpenReview](https://openreview.net/profile?id=~Rishi_Ahuja1) - [GitHub](https://github.com/RishiAhuja) - Email: rishia2220@gmail.com