LLM Applications | RAG | Document AI | Conversational and Voice AI | Agentic Workflows
Production AI that is accurate, auditable and fast, from first prototype to deployment.
Delhi NCR, India | Open to AI / GenAI Engineer roles (full-time, contract, remote)
AI Systems Engineer building production LLM applications: RAG pipelines, agentic workflows, real-time voice agents and OCR / document intelligence. Strong in Python and FastAPI, with a focus on guardrails, structured-output validation and production reliability. Previously at OnGraph Technologies, Head-Field Solutions and Paytm; currently working as a freelance AI Engineer delivering systems for clients.
| 95% | 87% | 9,271 | 500+ | 9,000+ |
|---|---|---|---|---|
| RAG retrieval precision | OCR benchmark accuracy | pages processed | AI evaluation tasks authored | TFUG community members |
Available for remote contract and freelance engagements. Click a service to email me about it.
| Service | What you get | Evidence |
|---|---|---|
| RAG and Knowledge Assistants | Retrieval pipelines on your documents and data: vector search, grounding, citations | 95% domain-specific retrieval precision in production |
| Document AI and OCR | Extraction and validation for scanned or handwritten documents, including fully offline and on-premise setups | 78% to 87% accuracy across 9,271 pages, zero silent errors |
| Voice and Conversational AI | Real-time voice agents and calling assistants with interruption handling and tool calling | Gemini Live, Telnyx, Ollama; public repos with 80+ stars |
| LLM Apps and Agentic Workflows | Multi-step agents with guardrails and audit trails; Dify and n8n automation into CRM and knowledge bases | PRD-to-QA pipeline cut test creation effort by 60% |
| AI Evaluation and Benchmarks | Test harnesses, reference solutions and structured-output validation to measure model and agent quality | 500+ agent-evaluation tasks authored |
| Backend and Deployment | FastAPI services, Docker, Kubernetes, CI/CD | Containerized services scaled in production |
Handwritten timesheets to Excel, fully offline | Document AI
- Problem: handwritten site timesheets had to be digitized for payroll with no cloud dependency and no silent mistakes.
- Solution: 3-stage OCR, extraction and validation pipeline (RapidOCR, FastAPI, multi-arch Docker) with benchmarking against Tesseract, EasyOCR and Qwen3-VL.
- Result: benchmark accuracy 78% to 87% across 9,271 pages, with zero silent errors. Part of an HRMS serving ~450 workers across 10+ sites.
Enterprise RAG at 95% retrieval precision | RAG and LLM Applications
- Problem: domain-specific questions needed accurate, low-latency answers from enterprise knowledge.
- Solution: RAG pipelines on vector databases with concurrent search, behind a stateful microservice conversational engine, with multi-model routing (OpenAI, Amazon Nova).
- Result: 95% domain-specific retrieval precision (OnGraph Technologies).
Real-time voice agent with barge-in | Conversational AI
- Problem: voice assistants feel slow and cannot be interrupted.
- Solution: bidirectional WebSocket streaming with an async audio pipeline, interrupt handling, concurrent sessions and tool calling.
- Result: sub-second conversational latency. Open source: ai-calling-assistant and NovaVoice AI.
| Project | Description | Stack |
|---|---|---|
| voice-agent-workshop (16 stars) | A fully local voice agent that hears you, talks back and acts (opens YouTube, searches, scrolls). No API keys, no cloud, no cost. | Whisper, Ollama, Kokoro, Playwright |
| voice_ai_agents_using-Gemini_live_api (47 stars) | NovaVoice AI: production-style Gemini Live voice agent with Google Calendar scheduling and a secure Python gateway. | Gemini Live, FastAPI, WebSockets |
| ai-calling-assistant (18 stars) | Real-time voice-to-voice calling assistant with function calling and interrupt handling. Zero backend. | Gemini Live API, Web Audio |
| novaflow-qa-agent (3 stars) | Turns a PRD into executable UI tests, runs them, and lets you talk to the report. Amazon Nova Hackathon 2026. | Amazon Nova, FastAPI, React |
| trueprice-cars24 | Variant-aware, confidence-scored used-car price estimator that closes the estimate-to-final-price gap. | Python, open-source MVP |
| echosphere-ps5-voice-support | Real-time multilingual voice support for e-commerce. Agora Conversational AI Hackathon. | Python, Agora |
| minimal-voice-ai-system | Streaming STT, LLM and TTS backend with barge-in, memory layers and hardware-aware tuning. | FastAPI, Whisper, Ollama, Docker |
Click a role to expand or collapse it.
AI Engineer (Freelance) | Jun 2026 to Present
- Built an offline 3-stage OCR, extraction and validation pipeline (RapidOCR, FastAPI, multi-arch Docker) that digitizes handwritten timesheets into Excel: 78% to 87% benchmark accuracy across 9,271 pages, zero silent errors.
- Architected an open-source HRMS for GCC construction operations: OCR digitization and payroll for ~450 workers across 10+ sites, validated against 49 requirements by automated test suites.
- Built a fully local, Dockerized OCR solution for certificate extraction and validation, plus an OCR engine benchmark lab (RapidOCR, Tesseract, EasyOCR, Qwen3-VL).
- Delivered a multi-agent document-processing platform (Prisma, PostgreSQL) and other client backends.
AI Trainer (freelance), Handshake and Turing | May 2025 to Present
- Authored 500+ technical evaluation tasks in Harbor-based environments and Terminal-Bench 2.0 setups: terminal workflows, coding tasks and multi-step AI-agent evaluations.
- Built reproducible task environments, reference solutions and tests to measure agent correctness.
- Reviewed hundreds of AI-generated and user-submitted tasks for technical accuracy, instruction adherence and benchmark compliance.
AI/ML Engineer, Head-Field Solutions | Noida | May 2026 to Jun 2026
- Built a production voice-to-voice agentic calling assistant with Ollama local LLMs, Telnyx and Claude Code.
AI Systems Engineer, OnGraph Technologies | Aug 2025 to May 2026
- Architected a low-latency conversational AI engine on stateful microservices; built enterprise RAG pipelines (vector DBs, concurrent search) at 95% domain-specific retrieval precision.
- Orchestrated multi-model workflows (OpenAI, Amazon Nova) with deterministic routing; built async agentic workflows in Dify and n8n connected to CRM and knowledge bases.
- Built FastAPI backends for low-latency LLM interoperability; shipped with Docker and Kubernetes.
Software Engineer Intern, Paytm | Jan 2025 to Aug 2025
- Built FastAPI backend services for high-throughput IVR systems.
Earlier roles and education
- Machine Learning Intern, Energy Venture (Jul 2024 to Oct 2024): IoT sensor data pipelines and computer-vision models for automated industrial inspection.
- Community Growth Manager, Where U Elevate (Oct 2023 to Apr 2024): market research and engagement work that doubled platform user adoption.
- B.Tech, Computer Science (AI and ML), NITRA Technical Campus, Ghaziabad (2021 to 2025).
Click to expand or collapse
| Area | Technologies |
|---|---|
| LLM and Agentic Systems | RAG, agentic architectures, function calling and tool use, LangChain, LangGraph, Dify, n8n, Gemini Live API, Amazon Nova and Sonic, prompt engineering |
| Document AI | RapidOCR, Tesseract, EasyOCR, Qwen3-VL, extraction and validation pipelines |
| Reliability and Evaluation | Structured-output validation, guardrails, traceability, automated testing, Selenium, benchmark design |
| Backend and Data | Python, FastAPI, Flask, REST APIs, SQL, PostgreSQL, Pinecone, Weaviate, pgvector, vector embeddings |
| Infrastructure | Docker, Kubernetes, Git, CI/CD, Google Cloud and Vertex AI, AWS, system design |
- Lead Organizer, TensorFlow User Group Ghaziabad: 9,000+ members; recognized by Google for Developers for community impact.
- Speaker and workshop host on Gemini, ADK, Vertex AI, agents and RAG.
- AWS Community Builder, AI Engineering (2026); AWS Machine Learning Summer School Finalist.
- ICPC Regionalist (2023 to 2024).
Open to remote contract, freelance and full-time AI / GenAI engineering roles.
How an engagement typically runs: 1. short call to scope the problem and success metric 2. working prototype measured on your data 3. production hardening and deployment 4. handover with documentation.
tripathiakaash@gmail.com | Medium