Vol. XXVIIPrice: Free

The Sidhardh Times

Extra! Extra! Read all about it

AI ENGINEER
ENTERS THE BUILDING

Vol. XXVII · No. 1 “All the intelligence that’s fit to print” Price: Free

The Sidhardh Times

Kottayam, India • 2026 • Late City Edition • Weather: bright, with scattered gradients

Exclusive

Student Builds Machines
That Learn to See Risk

By our Technology Correspondent

A final-year student at VIT-AP has spent the better part of two years teaching graphs to reason about danger. The result, submitted this season to IEEE Transactions on Intelligent Transportation Systems, stitches together three ideas nobody had bolted together before.

“The interesting part isn’t the architecture,” Sidhardh says. “It’s the ablation — the experiment that proves which piece actually earned its place.”

A second front has opened in low-resource languages, where a new benchmark hunts for fraud in the code-mixed SMS traffic of five Indian tongues.

Continued in Research, p. 2

Engraving: a figure in a long coat and hat pauses beneath an iron footbridge on the embankment at night — **ON THE EMBANKMENT.** Our correspondent takes the night air beneath the footbridge, thinking about ablations.

“My neural networks might be artificial,
but the bugs I write are entirely real.”

Anyone can post a number. The work is the dataset, the ablation that settles the argument — and then, unfashionably, deployment.

Stop Press

PyTorch✦Graph Neural Networks✦ Mamba State Space Models✦Quantized LLMs✦ Uplift Modelling✦Explainable AI✦ FastAPI✦Adversarial Robustness✦ PyTorch✦Graph Neural Networks✦ Mamba State Space Models✦Quantized LLMs✦ Uplift Modelling✦Explainable AI✦ FastAPI✦Adversarial Robustness✦

Page 2

Research & Publications

Graph learning for road-safety risk; quantized language models for fraud in code-mixed Indian languages.

Illustration: a lone figure at a cluttered workstation of monitors and machinery, lit blue — **THE WORKBENCH.** Where the graphs get built, the ablations get run, and the arguments get settled. Hours: unsociable.

Submitted · IEEE T-ITS

HeteroRisk

HGT + Mamba SSM + Zero-Inflated Tweedie for crash-risk prediction

A Heterogeneous Graph Transformer over four node types and five typed edge relations, a Mamba selective state space model for linear-complexity temporal memory, and a Zero-Inflated Tweedie output head — the first known integration of all three in a single model.

Built on a road-safety graph assembled from 175,000 California crash records and 27,615 OpenStreetMap road segments, and validated by component ablation and gradient-based counterfactual policy explanations.

MAE	F1	vs GAT+GRU
0.0261	1.000	−58%

PyTorch Geometric · Mamba · Graph Transformers · Tweedie

Capstone · in progress

IndoSmish

Code-mixed SMS phishing detection with quantized LLMs

The first multilingual code-mixed smishing benchmark for Indian languages: roughly 3,000 examples spanning Hinglish, Manglish, Tanglish, Bengali and English, covering the UPI scam categories that dominate Indian mobile fraud.

Four-bit quantized models — Phi-3-mini, Gemma-2, Qwen-2.5 — are benchmarked against classical baselines and the Indic encoders MuRIL and IndicBERT, with an adversarial-robustness study and edge-deployment profiling.

Examples	Languages	Precision
~3,000	5	4-bit

Target venues: ACSAC 2026 · KDD Applied Track 2026

Page 3

Dispatches from the Workbench

Selected projects, most recent first.

Noir illustration: a figure in a hat and coat solving a Rubik's cube against a city skyline — **ON THE BEAT.** Every project starts the same way — something scrambled, and a suspicion that it needn’t stay that way.

Overhead illustration of a desk: laptop, notes, coffee and a pair of boots up on the table — **THE DESK, 2 A.M.** Coffee cold, notes everywhere, boots up. This is where most of the arguing gets done.

Top 10 · Epsilon TeXpedition

Foresight

A counterfactual “what-if” simulator for marketing journeys. An S/T-learner uplift engine behind a FastAPI service, with a React and Recharts front end and Gemini integration. Recognised as a Top-10 finalist at Epsilon’s flagship campus hackathon in Bangalore.

See the live demo ⟶

Biometric EHR

Tamper-proof health records: hashes anchored on-chain, encrypted payloads on IPFS, and biometric authentication that stores only irreversible templates.

Solidity · IPFS · AES-256

ClipBridge

Zero-install clipboard sharing across devices via a four-digit code. 42 ms median delivery, 118 ms at the 95th percentile, 87 per cent statement coverage.

WebSockets · PostgreSQL · Cypress

Disease Prediction

132 symptoms mapped to 41 diseases across 4,920 samples, with six classifiers compared under cross-validation and hyperparameter tuning.

scikit-learn · XGBoost · Pandas

Page 4

The Man Himself

Sidhardh S is a final-year Integrated M.Tech (Software Engineering) student at Vellore Institute of Technology, Andhra Pradesh, and lives in Kottayam. His work runs from heterogeneous graph neural networks and state space models at one end to quantized language models for low-resource, code-mixed languages at the other.

He is currently seeking an AI/ML internship — preferably one where research-grade work is expected to ship.

Agentic AI · DeepLearning.AI Spec-Driven Development · DeepLearning.AI

Illustration: a figure pressing a loaded barbell on a bench — **OFF THE CLOCK.** Progressive overload — the only other place an ablation study makes sense.

2022 — 2027
Integrated M.Tech, Computer Science & Software Engineering

Vellore Institute of Technology, Andhra Pradesh
Jan 2024
Artificial Intelligence Intern — FlyRank AI

Model integration, data pipelines and feature development.
Dec 2023 — Jan 2024
Artificial Intelligence Intern — CodSoft

Image-recognition pipeline in Python with OpenCV and pre-trained CNNs.

The Toolkit

Languages: Python · JavaScript · SQL · Java · C/C++
ML & DL: PyTorch · PyTorch Geometric · TensorFlow · scikit-learn · XGBoost
GenAI: Hugging Face · 4-bit quantization · RAG · MuRIL / IndicBERT
Backend: FastAPI · Node · Express · WebSockets · REST
Data: PostgreSQL · MongoDB · AWS · Docker · Git
Concepts: Graph neural nets · State space models · Uplift modelling · Explainable AI

Page 5

Classifieds

Situations wanted. Replies by post, wire or electronic mail.

Electronic Mail sidhardh215@gmail.com Fastest reply ⟶ Code Repository codewithcid On GitHub ⟶ Professional Notices Sidhardh S On LinkedIn ⟶ Amusements The Interactive World Take the wheel ⟶

And finally —

the proprietor leaves the building

Student Builds MachinesThat Learn to See Risk

“My neural networks might be artificial,but the bugs I write are entirely real.”