I design and build production-grade Generative AI applications, full-duplex real-time voice agents, warehouse-native RAG architectures, and multi-agent orchestration pipelines on Google Cloud.
flowchart TD
Portfolio["Shubham's Generative AI Engineering Portfolio"]
Portfolio --> V["1. Flagship Industry Agents\n(Fintech, Retail & AgriTech)"]
Portfolio --> R["2. Real-Time Voice & Speech\n(Gemini Live WebSockets & Chirp 3)"]
Portfolio --> E["3. Enterprise RAG Pipelines\n(BigQuery Vector Search & Vertex AI)"]
Portfolio --> A["4. Multi-Agent Orchestration\n(Google Agent Development Kit)"]
V --> V1["smb-loan-visual-underwriter"]
V --> V2["omnichannel-retail-stylist"]
V --> V3["agritech-crop-weather-copilot"]
R --> R1["argus-gemini-live-debate-agent"]
R --> R2["pipecat-chirp3-voice-cloning"]
R --> R3["websocket-pcm-audio-streamer"]
R --> R4["longform-multilingual-tts"]
E --> E1["bigquery-vector-search-rag"]
E --> E2["first-principles-vertex-rag"]
A --> A1["sequential-newsroom-orchestrator"]
A --> A2["financial-valuation-tool-agent"]
| Repository | Domain | Architecture Highlights | Stack |
|---|---|---|---|
smb-loan-visual-underwriter |
Fintech / Credit Risk | Evaluates physical storefront & workshop photos using Gemini 2.5 Vision to estimate inventory value, asset collateral, and a 0–100 business health score for thin-file micro-merchants. |
Python, Streamlit, Gemini 2.5 Flash |
omnichannel-retail-stylist |
Retail / E-Commerce | Combines Gemini 2.5 visual reasoning with Imagen 3 (imagen-3.0-generate-002) for visual wardrobe matching, virtual try-on generation, and localized catalog translation. |
Python, Vertex AI, Imagen 3, Streamlit |
agritech-crop-weather-copilot |
AgriTech / Sustainability | Multimodal leaf pathology & soil diagnostics fused with live Google Maps Weather API telemetry and multilingual voice synthesis for farmers. | Python, Gemini 2.5, Google Maps API |
| Repository | Pattern | Architecture Highlights | Stack |
|---|---|---|---|
argus-gemini-live-debate-agent |
Full-Duplex Voice Agent | Sub-second interruptible voice debate & clinical communication simulator with custom React 19 AudioWorklet, Google Search grounding, and automated rubric evaluation. |
React 19, TypeScript, FastAPI, Gemini Live API |
pipecat-chirp3-voice-cloning |
Instant Voice Cloning | Captures a live enrollment phrase and converses back in the speaker's cloned voice using Pipecat and Cloud TTS Chirp 3 Instant Custom Voice. | Python, Pipecat, FastAPI, React, Chirp 3 HD |
websocket-pcm-audio-streamer |
Raw PCM Duplex Stream | Hardware-level asyncio + PyAudio client streaming 16kHz mic audio + live OpenCV frames and playing back 24kHz PCM audio with VAD interruption handling. |
Python, Asyncio, PyAudio, OpenCV, WebSockets |
longform-multilingual-tts |
Neural Audio Synthesis | Unified engine for two-host podcast generation (MultiSpeakerVoiceConfig), Chirp 3 HD voices, and asynchronous Cloud TTS Long Audio synthesis to GCS. |
Python, Google GenAI SDK, Cloud TTS v1 |
| Repository | Vector Engine | Architecture Highlights | Stack |
|---|---|---|---|
bigquery-vector-search-rag |
BigQuery VECTOR_SEARCH |
Serverless PDF ingestion, sliding-window chunking, Vertex AI text-embedding-004 indexing, and SQL-native VECTOR_SEARCH cosine retrieval inside BigQuery. |
Python, BigQuery, Vertex AI Embeddings, Gemini 2.5 |
first-principles-vertex-rag |
In-Memory NumPy Index | Transparent, zero-framework implementation of document chunking, 768-d vector indexing, exact NumPy cosine similarity `(u · v) / ( |
| Repository | Orchestration Pattern | Architecture Highlights | Stack |
|---|---|---|---|
sequential-newsroom-orchestrator |
SequentialAgent Pipeline |
Chains a Search-grounded Scout Agent, Writer Agent, and Editor Agent across shared InMemoryRunner session state to produce fact-checked briefings. |
Python, Google ADK (google-adk), Gemini 2.5 Flash |
financial-valuation-tool-agent |
Deterministic Tool Calling | Equips an ADK Agent with strongly-typed Python financial tools (CAGR, P/E implied equity valuation, sector multiple benchmarking) to eliminate LLM math hallucinations. |
Python, Google ADK (google-adk), Gemini 2.5 Flash |