Chandan Kumar

• Building the core agent engine behind every call on OSVI's voice AI platform: ASR, LLM reasoning, TTS and telephony.
• Built voice-to-voice calling on Gemini Live and OpenAI Realtime, cutting first-response latency, with fallback to the STT-LLM-TTS pipeline so calls survive model outages.

• Built a fintech KYC system that classifies documents, extracts fields with OCR and matches faces to ID photos.
• Hardened the pipeline against spoofed scans and low-light selfies, reducing false rejects within product latency SLAs.

• Owned the model lifecycle: data preparation, training, evaluation and low-latency inference APIs for product teams.
• Built evaluation harnesses and regression baselines so model upgrades could be compared, debugged and rolled out safely.

• Built a ResNet-50 plant-disease classifier (93% accuracy, 87K+ images) served via FastAPI with batched inference, plus a recommender with NLP sentiment analysis over agricultural news.

• Built a YOLOv5 and PaddleOCR pipeline that digitised 500+ legacy engineering drawings; denoising, adaptive thresholding and deskewing improved robustness by 15%, and my CRNN vs Tesseract vs PaddleOCR benchmark set the team's OCR stack.

Python, FastAPI, Presidio, spaCy NER, MCP

Masks PII with reversible vault tokens before prompts reach LLMs; shipped as a library, FastAPI middleware, CLI and MCP server for Claude Code and Cursor. Drew seed-funding interest from Apertu Capital.

llama.cpp, Metal, FastAPI, React

Importance-matrix-calibrated IQ1_S quantization shrinks Llama-3.1-8B from 16.1 GB to 2.19 GB (7.4x), running at ~51 tokens/s in 3.1 GB of RAM, with a live side-by-side benchmark playground.

Hugging Face, LoRA/PEFT, 4-bit quantization, LangChain, ChromaDB

Fine-tuned Mistral-7B on 10K+ Q&A pairs with LoRA, cutting perplexity by 35% and the 4-bit model from 14 GB to 4 GB at under 2% accuracy loss; the RAG pipeline reached 91% retrieval accuracy.