Divyansh
Teja Edla

India has 63 million small businesses. Most are invisible online — not because they don't matter, but because no one built them the right tool. A shop owner has receipts, photos, and price lists — but no website, no structured data, no digital presence.

• Designed a multi-agent LangGraph pipeline where each agent handles one stage of document understanding
• File Discovery Agent scans uploaded ZIPs and routes documents by type (PDF, DOCX, images, videos)
• Specialized agents extract tables, parse text, and analyze media in parallel
• Schema Mapping Agent structures extracted data into a unified business schema
• Validation Agent checks for consistency and completeness before final output

In federated learning across hospitals, larger institutions dominate model performance. Standard FedAvg averages weights proportionally — so a hospital with 5,000 scans drowns out a rural clinic with 200. The patients who need AI most get the worst model.

• Replaced FedAvg weight aggregation with Quantum-inspired Particle Swarm Optimization (QPSO)
• QPSO explores the weight space stochastically, finding aggregation points that reduce performance variance across clients
• Built 3D Attention U-Net for volumetric brain tumor segmentation on BraTS 2021 (4 MRI modalities)
• Compared FedAvg vs FedProx vs QPSO-FL across 3 simulated hospital nodes with Non-IID data splits
• Added LSTM-based tumor progression prediction as a downstream clinical module

Telugu is spoken by 80 million people. Most AI produces 'translated English' — grammatically passable but culturally hollow. Quality Telugu storytelling requires a system that understands the language's idioms, pacing, and literary tradition, not just its grammar.

• Built a multi-stage cognitive loop: Plan → Draft → Critique → Polish
• Planning Agent retrieves culturally relevant context from a 10,000+ story corpus via semantic search
• Drafting Agent generates Telugu text grounded in retrieved context using GPT-OSS-120B
• Validator Agent acts as a senior editor — rejects drafts using passive voice, lazy descriptions, or cultural inaccuracies
• Multiple revision cycles until the Validator passes the output — quality enforced architecturally, not prompted

Classic Telugu literature — stories, poems, serials from the iconic Chandamama magazine — exists in physical archives slowly disappearing. No structured digital corpus exists. Without one, this cultural heritage is one generation from being lost.

• Digitized and indexed 10,000+ stories from Chandamama archives into a vector database
• Built semantic search over the corpus using Alibaba-NLP/GTE Multilingual embeddings
• Created a 'Council of Storytellers' — 5 LLMs generating from the same prompt, judged on literary quality
• Evaluation criteria: Show-Don't-Tell, pacing, native idiom usage, cultural authenticity
• Added author analysis, character mapping, and location-based story graphs as analytical tools

Telugu poetry has a rich tradition of prosodic meters (Chandassu) — Vritta, Jati, Upajati padyam — that modern NLP tools completely ignore. No existing tool can classify Telugu verse meters or analyze syllable weights (Laghu/Guru).

• Studied Telugu prosody rules from classical grammar texts — no training data needed
• Built a rule-based engine that decomposes Telugu text into syllable units
• Each syllable classified as Laghu (light) or Guru (heavy) based on vowel length and consonant clusters
• Pattern matched against known Chandassu meters from classical Telugu poetry treatises
• Planning to release as a standalone pip-installable Python library

Credit card fraud detection models achieve high accuracy, but banks and analysts can't deploy what they can't explain. A flagged transaction with no explanation is useless — analysts need to know which specific features drove the prediction to trust and act on it.

• Trained multiple classifiers: XGBoost, Random Forest, Logistic Regression, KNN, and a Hybrid Stacking ensemble
• Applied SMOTE to handle extreme class imbalance (0.17% fraud in 284,807 transactions)
• Integrated SHAP (SHapley Additive exPlanations) for per-feature attribution on every prediction
• Built interactive Streamlit dashboard showing global feature importance and per-transaction explanations
• Designed the UI so analysts see not just 'fraud/not fraud' but exactly which features drove the decision