sushant nemade

Master's thesis adapting the dual-head MCI architecture from Change-Agent / LEVIR-MCI (Liu et al., IEEE TGRS 2024) with bidirectional temporal-reversal (A ↔ B) augmentation across 10,077 original and 15,115 augmented pairs (468- to 714-token vocabularies). Reorganized into an installable src/urban_change package without legacy MMCV coupling, featuring safe checkpoint loading, token-identity vocabulary transplantation, staged training with full RNG/scaler resume, and a shared predictor across CLI, Streamlit and bounded FastAPI endpoints.

Verified by 31 passing pytest/integration checks, numerical equivalence within 3.81e-6 against the legacy graph, and SHA-256 verified CPU inference on real chronological and reversed LEVIR-MCI test pairs. Full-dataset benchmark reproduction, GPU/AMP training and mixed upstream licensing (MIT Change-Agent/COCO-caption plus NVIDIA non-commercial SegFormer terms) remain documented boundaries.

Seven autonomous agents run a perceive, remember, reason and act loop over isolated player views, with bounded retries that fail closed on invalid schema or domain targets. Includes 27 deterministic offline tests and a batch simulator for ELO and role win-rate comparison. Session state is in-memory and the default mock provider is a deterministic heuristic baseline.

Hand-written 8→12→2 feedforward policy network, 7-ray obstacle perception, bicycle-style physics and background evolution across training, manual drive and champion autopilot modes. Runs entirely client-side with localStorage policy checkpointing and JSON export/import.

Real phone, SMS, database, AI and escalation flows remain unverified without credentials. The preview is not an active customer-call service.

The simulator sends no real WhatsApp messages. Deduplication is in-memory; durable sessions, booking and structured OpenAI classification are not implemented.

The public dashboard uses fictional data. Complete durable OAuth is not implemented, and scheduled delivery requires explicit configuration. It is not an active personal inbox assistant.

The public demo extracts actions with local rules. Real transcription requires configured OpenAI access; authentication and a permanent server-backed note archive are not configured.

Adapted from Hema Kalyan Murapaka's GPLv3 TravelBrain. Twenty-seven Python tests, 27 cross-runtime planning combinations, 14 browser cases, installed-package checks and audits pass. Live Open-Meteo weather was verified; Groq inference remains unverified. Prices are illustrative; no flight/hotel availability, booking or accessibility guarantee is provided.

Adapted from Hema Kalyan Murapaka's MIT-licensed BlogBoard. Offline and desktop/mobile tests pass; the article preview is a synthetic fixture. Live Groq, search and R2 integrations remain unverified.

The public page uses fictional sample data. Private routes use a single admin token; Stripe, Shopify and QuickBooks adapters are future work, not connected integrations.

Original code inspired by GitPulse's documented concept; upstream source was not imported because its license file could not be verified. One public REST sync was tested. The public dashboard uses fictional data; private API hosting remains unverified.

Adapted from Hema Kalyan Murapaka's GPLv3 MarketInsight. Complete sample/API workflows, Linux/Windows CI, package checks and a non-root container pass. Real Yahoo quote/history was verified; paid AI inference remains unverified. The public workspace uses fictional data and provides research, not investment advice.

Adapted from Hema Kalyan's AGPLv3 Doclify. Offline workflows, installed packages and Linux/Windows CI pass. Offline output is structural, not AI inference; live Groq generation remains unverified. This is a local CLI, not a hosted AI service.

Original code. Go build info, Rust cargo-auditable, npm/ASAR, NuGet, Python metadata and Maven pom.properties inventories (including nested Spring Boot and WAR archives), Electron fuse and webPreferences rules and the Authenticode signer of Windows executables; vulnerability lookup is opt-in and sends only package name, version and ecosystem. The update check gates on new vulnerabilities, weakened fuses, string-literal network endpoints, new native modules, removed signatures and a changed publisher. v0.3.0 parsers are checked in CI against rust-audit-info on Linux, Windows and macOS, Maven Central and PyPI artifacts (Log4Shell inside a nested JAR) and Windows Get-AuthenticodeSignature; published on PyPI and the GitHub Marketplace with PEP 740 provenance attestations. Findings are version-level evidence, not reachability analysis.

Original code. Test-persona matching including hashed, Base64 and gzip-compressed forms, WebSocket messages with Socket.IO and SignalR framing, checksum-verified IBAN, Steuer-ID and KVNR detectors, health-data fields and a GDPR-oriented data-flow map; third-party eTLD+1 connections are always inventoried, with opt-in connection gating. Reports never contain raw values and the redacted HAR export is self-checked. A Playwright Test (Node) fixture records WebSocket messages, and CI requires identical findings from Python and Node recordings of a fictional demo shop with real Chromium. Technical evidence for privacy reviews, not legal advice.

Social-platform ingestion, model-based clustering, persistent storage and Notion export are not connected. Analysis starts from directly supplied text.

Profiles are fictional and requests are local-only. Account creation, embeddings, pgvector and messaging are not configured.

Image understanding and route optimization are not active. Grouping uses entered city labels, not geocoded distances; no travel-time or booking guarantees.

There is no shared server database, authenticated administrator or automatic social posting. Local edits do not publish to Twitter/X or LinkedIn.

Editing affects only the current browser. This is not a shared, authenticated CMS; server identity, a database and production analytics controls are not implemented.