2025 – 2026·Final Year Project·YOLOCLIPFaceNetFastAPIReactMongoDBMinIO
Multimodal CV+NLP platform for real-time CCTV threat detection. YOLO detection module identifies weapons, intrusions, and behavioral anomalies. CLIP-based vision-language pipeline enables natural language queries against surveillance footage — operators may search "person near entrance" and receive indexed clips. FaceNet identity tracking associates recurring individuals across camera feeds. NLP incident reporting generates structured threat summaries with keyframes, risk indicators, and searchable metadata. Backend: MongoDB for event logs, MinIO for video clip storage, RTSP stream integration.
2025·LangChainCrewAIRAGSecurity
CLI-based multi-tenant RAG architecture with tenant isolation, ACL enforcement, PII masking, and prompt injection detection. Buffer and summary memory modes implemented with auditable logging. Strict policy guards enforced at execution level. Passed red-team evaluations with full audit trail.
2025·ReActReflexionCoTLangGraph
Heterogeneous autonomous agent system simulating medic and drone coordination across earthquake scenarios. Benchmarked ReAct, Reflexion, and Chain-of-Thought reasoning frameworks under realistic hardware constraints (battery, debris). GUI metrics dashboard built for cooperation analysis and multi-agent strategy evaluation.
2025·CrewAILangGraphOpenSky APIMCP
Live flight anomaly detection with real-time data via OpenSky API. CrewAI/LangGraph orchestration with MCP tool integration. Snapshot caching and graceful API fallback implemented for operational resilience. Delivered comprehensive technical report with full evaluation.
2026·SemgrepAST AnalysisLangGraphNext.jsSupabase
Traces real multi-hop taint paths through an uploaded codebase — Semgrep anchors source/sink points, a custom Python AST tool builds the cross-function call graph, and an LLM reasons hop-by-hop over the real source to tell a genuine sanitizer from a cosmetic no-op. Confirmed findings get an auto-drafted PoC and fix diff. Scans run as on-demand GitHub Actions jobs against uploaded repos.
2026·LLM AgentsNext.jsSupabaseGroq
Ingests multiple witness depositions from the same case, builds per-witness timelines, and cross-references statements to flag contradictions backed by exact page/line citations. A deterministic regex parser anchors every citation to the real transcript so the LLM can never invent one; a separate cross-reference pass tells a genuine contradiction apart from statements that only sound alike.
2026·Reactffmpeg.wasmpdf-libVercel
91 real tools across PDF, image, video, audio, developer, and text utilities — every one runs entirely client-side, no file ever leaves the browser. The one exception is LaTeX compilation, sandboxed behind a rate-limited serverless function since no maintained WASM LaTeX compiler exists. Real WCAG contrast math, real dominant-color extraction, real EPUB spec compliance — not stubs standing in for the real thing.
2026·Kokoro TTSffmpeg.wasmWeb WorkerReact
Paste a script, drop in a video, get back a narrated video with burned-in captions — entirely local, no backend, no API keys. Kokoro TTS runs in-browser via WASM inside a Web Worker so synthesis never freezes the UI; timestamped script sections anchor narration to real timing instead of drifting; ffmpeg.wasm burns word-wrapped captions and swaps the audio track client-side.
2026·Gemini VisionPaddleOCRWCAGSaliency ModelConvex
Upload a UI screenshot or paste a URL, get back a scored audit across layout, accessibility, copy, and attention — each finding cited to something checkable, not an LLM's vibes. Three of four lenses are deterministic code (real WCAG contrast math, OCR-measured tap targets, an MSI-Net saliency model's attention-vs-CTA overlap); only qualitative layout critique is a model call. Exactly two AI calls per audit, both bounded.
2026·Manifest V3GroqTavily$0 Cost
Select text on any page, right-click "Fact-check this," and an agent searches the web and shows a sourced verdict badge inline. Genuinely contested claims get both perspectives with their own sources instead of a forced single answer. No backend, no build step — one Tavily search plus one Groq call per check, entirely client-side.
2026·Manifest V3Groq gpt-oss-120b$0 Cost
Highlight a sensational headline, right-click, and get it rewritten in plain language with its manipulation tactic named and scored 0–100. Six files of vanilla JavaScript, one LLM call per decode, strict JSON output at temperature 0 — no backend, no framework, no dependencies.
2026·Manifest V3Groq Vision$0 Cost
Removes AI-generated posts and reels from Instagram feeds automatically. Tier 1 catches Instagram's own "Made with AI" labels for free with no API call; opt-in Tier 2 sends unlabeled thumbnails to a free Groq vision model. No backend, no build step, no dependencies — a MutationObserver and one content script.
PythonOllamaMediaPipescikit-learnOpenCV
Story2Audio
[GitHub ↗]
— LLM-driven emotionally expressive audio synthesis pipeline using F5-TTS and gRPC microservice architecture with Gradio frontend.
MPCount
[GitHub ↗]
— Retail inventory counting via deep learning on SKU-110K dataset using crowd-counting methodology.
Agentic OCR
[GitHub ↗]
— VLM-powered agent pipeline for structured document extraction.
Flight Delay Prediction
— 96% accuracy ML pipeline for aviation delay classification.
Virtual Painter
[GitHub ↗]
— Real-time gesture-controlled drawing tool using MediaPipe hand tracking and OpenCV.
FAQ Chatbot
[GitHub ↗]
— NLP-based FAQ retrieval chatbot using semantic similarity and transformer embeddings.
GANs & VAEs
[GitHub ↗]
— Generative modeling pipeline covering signature synthesis and image generation via GANs and Variational Autoencoders.
