Birat Gautam

Experience & Education

1. N
   NepaWorksAI/ML Engineer

   Jul 2026 – Present#e6d2b74

   • Architected an end-to-end multimodal lead-qualification platform processing 300,000+ commercial retail businesses and 65,000+ store photos across distributed Playwright scrapers, local LLM inference, and human review, achieving 99.95% reliability.
   • Serve Gemma 4 on-prem on an RTX 5090 via a Blackwell-tuned llama.cpp build (300-400 classifications/hour).
   • Achieved 99.0% to 100% lead precision (0 false leads on held-out benchmarks) and 88% product-line accuracy at $0 marginal API cost (matching 97% of cloud Gemini Pro/Flash reference accuracy) by upgrading on-prem inference from 12B to Gemma 4 31B Dense.
   • Traced classification errors to missing image detail rather than the prompt by testing one change at a time; raised image resolution to 2048px (1,120 vision tokens/image, up to 15 photos / about 18.5K tokens per store), which fixed blurred shelf-level products.
   • Eliminated VLM confidence miscalibration (where 90% of false negatives reported 1.00 confidence) by replacing score thresholds with JSON-schema grammar citations requiring exact photo indexing, explicit verification flags, and a deterministic 2% live audit sample.
   • Built the async FastAPI and PostgreSQL control plane for distributed workers: lease-based claiming, transaction-scoped advisory-lock leader election, PgBouncer-safe; 35 migrations and 2,400+ automated tests.
   • Optimized the pipeline: cut export latency 24x (448 ms to 18 ms) via batched queries, reduced host server RAM 75% (36 GB to 9 GB) via sliding-window checkpoint caps (--ctx-checkpoints 4), and cut CDN download bandwidth by about 90%.
2. S
   Synapse TechnologiesJr. AI/ML Engineer

   May 2025 – Jun 2026#a3f91c2

   • Engineered a multilingual NLP pipeline for English, Nepali, and Romanized Nepali queries using transformer-based intent classification and script-aware NER; reduced inference latency from ~10s to ~30ms via ONNX optimization and INT8 quantization.
   • Developed and productionized the KanoonBox RAG pipeline using ChromaDB vector search and NepSearch-based legal grounding; optimized intent-routing architecture to reduce reasoning latency from ~5s to ~25ms while maintaining response quality.
   • Deployed Nepali document OCR using PaddleOCR-VL-0.9B via llama.cpp within 2 GB VRAM (7s/page); integrated Qwen2.5-VL-3B-Instruct-GGUF into Synapse Intelligence for enterprise document understanding and knowledge delivery.
   • Built Nepal's first structured Nepali ASR corpus from YouTube audio using VAD, NFC normalization, and Gemini validation; LoRA fine-tuned Whisper-large-v3-turbo on resource-limited hardware (WER: 33, WandB monitored).
   • Benchmarked lightweight Gemma 4 variants on CPU infrastructure using Ollama, llama.cpp, and GGUF quantization formats for on-device LLM deployment feasibility.
3. B
   Birmingham City UniversityBSc (Hons) CS with AI

   2023 – Present#7b2d441

   • Currently Third Semester Student.
   • Spearheaded a team of five to build a face-recognition attendance web app.
   • Completed First Year with First Class Honours.
4. D
   Deerwalk Compware Ltd.Backend Developer Intern

   Jun 2023 – Aug 2023#4e8c910

   • Designed and developed an Attendance Management System using Django.
   • Implemented Excel report generation for detailed attendance records.
   • Built an Admin Panel with full CRUD operations.
5. U
   Uniglobe SS/CollegeHigh School Graduate

   2020 – 2022#1f2a398

   • Graduated with a 3.78 / 4 GPA.
   • Awarded Student of the Year — UGSS Achievers Award 2022.