Shyam Sunder Chiliveri

Problem

Transfer-pricing files mix OECD guidelines, screening of comparables, and a margin test. The question is which paragraph was actually retrieved, and whether the tested party sits inside the range.

Approach

A router sends each question to guideline RAG, a knowledge graph, company screening, a TNMM margin test and a FAR characterisation. A critic checks that every citation was retrieved. The worked example is the fictional hardening group Härtetechnik AG.

Result

Top-3 citation accuracy 45/45 on the published question set. Company screening F1 about 94%. Public page plus an interactive Docker desk.

45/45 citation accuracy

Problem

Embedded CAN code can hide integer wrap and out-of-bounds writes that a reviewer will miss in a long C file. The desk has to show a real finding, a CWE, and a suggested bound, without claiming to be an OEM cybersecurity programme.

Approach

Same shape as the materials case desk: router chooses specialists, tools return facts, critic last. Target is Open-SAE-J1939 (about 33 .c files). The public page replays the first SAST scan so a recruiter can open the board without Docker.

Result

First builtin SAST run: 4 raw hits kept by the critic, including TP.DT index wrap (CWE-190) and a write past MAX_TP_DT (CWE-787). Live GitHub Pages desk, Docker rescan on a clone.

4 critic-kept findings

Problem

What makes bees busy? Hive flow, temperature, humidity and weight sit in messy CSVs, in German local time, with stacked series and sensor sign errors. They need to meet hourly weather before anyone can answer the business question.

Approach

Source CSVs copy into bronze, archive with timestamps, then Synapse notebooks validate and write Parquet to silver. A second notebook joins weather and builds hourly, daily and insight tables in gold. If a quality check fails, the pipeline stops and files stay in processing.

Result

Full Data Factory run succeeded (14 activities). Data from January 2017 to May 2019, hive in Bad Schwartau. ARM template of the factory is in the repo.

2.5M flow rows

Problem

Materials engineers need a traceable failure analysis for hardened steel axles: classify SEM fracture images, retrieve the relevant ISO/DIN standards, and estimate fracture risk from plant motor-load data on 4Cr13 axle steel.

Approach

A router dispatches each case to 6 specialist agents — standards RAG via FAISS, knowledge graph, fracture ML, SEM vision, digital twin — and a critic agent reviews the answer. Every case returns one traceable answer: cause, process, next inspection.

Result

Deployed via Docker on AWS ECS Fargate with CloudFront and Secrets Manager (Frankfurt), pytest CI/CD and a Streamlit UI — publicly available as a live app.

Router + 6 specialist agents

Problem

Materials characterisation needs three things at once: predicting ductility/brittleness, segmenting fracture surfaces in SEM images, and answering questions across thousands of ISO/DIN standards.

Approach

An XGBoost ductility/brittleness predictor, a U-Net for SEM fracture segmentation, and a RAG agent with full tool-calling over 13,400+ ISO/DIN standards — backed by a Qdrant vector DB, SHAP explainability, a Streamlit UI and full pytest CI/CD.

Result

R² = 0.90 on the predictor, 95% segmentation accuracy, 385K training samples and 13,400+ documents indexed.

R² = 0.90

Problem

Reliably separate fake from real news across a corpus of 44,000 news articles.

Approach

A full NLP pipeline comparing classical ML (TF-IDF + SVM, Logistic Regression) against fine-tuned BERT and RoBERTa transformers, trained on an RTX 3060 GPU and evaluated with a Bonferroni-corrected framework.

Result

Fine-tuned BERT reached 98.62% accuracy, against 94.04% for the best classical model (SVM).

98.62% BERT accuracy

Problem

An outdoor sports supplier operating across 21 markets needed a self-service view for the CEO covering revenue, margin, seasonality, European expansion, retailer performance and marketing KPIs.

Approach

An end-to-end pipeline — Google Sheets → BigQuery (14 SQL CTE queries over 149,257 transactions) → Google Data Studio — delivered as a dashboard with 5 stakeholder pages.

Result

€1.25B of revenue analysed across 149K transactions, 21 markets and 289 retailers.

€1.25B revenue analysed

Problem

Classify a large image corpus at production quality, then serve the winning architecture behind a real inference endpoint — not a notebook demo.

Approach

Trained four CNNs (MobileNetV2, ResNet50, EfficientNetB0, DenseNet121) with transfer learning on 60,000 images, compared them, and shipped the best model through TensorFlow Serving, Docker, GCP and a CI/CD pipeline with automated evaluation logging.

Result

A production deep-learning stack on GCP — 4 architectures trained, best model deployed with full MLOps.

60K images