CV

Probabilistic ML, simulation, and MLOps for public procurement · built solo

• Probabilistic ML & MLOps | Built TabICLv2 and AutoGluon models to predict bidder counts and winning prices, and simulated competitor bids with Monte Carlo methods. Compared TabPFN, TabM, and RealMLP on chronological splits to select regional models. Automated notice collection, training, calibration, GPU batch inference, and deployment on AWS with Airflow, S3, and EC2, improving the bid-award KPI by 30% over the previous quantile model and reducing processing time from 136 to 25 seconds.

Probabilistic ML, NLP, and RAG for public procurement · built solo

• PQ Bid Prediction | Modeled multimodal bidding behavior across eight competitive contexts. Fixed collapsed R2CCP prediction intervals using entropy regularization and per-bin conformal thresholds, reaching 90.73% coverage in chronological evaluation. Built the data, training, 500,000-sample Monte Carlo batch, and web service in two months, improving the group's PQ bid-award KPI by 35%.
• Procurement NLP | Fine-tuned RoBERTa-large with LoRA to classify bid eligibility and business categories from notice titles. Used domain-specific weak supervision and teacher–student distillation to address limited labels. Deployed INT8 ONNX models through FastAPI on CPUs, achieving eligibility Macro-F1 of 0.9639 and 50 ms inference per notice.
• Document Retrieval & RAG | Indexed OCR text, construction standards, and unit-cost data in FAISS. Connected document retrieval and cost calculation through LangGraph to generate draft quantity takeoff sheets and official correspondence.

Document, CAD, and surveying workflow automation

• Developed format-specific OCR, AutoCAD add-ins, and ERP integrations, reducing document work from over four hours to about 30 minutes.

Product image processing and data automation with OpenCV

• Used OpenCV to extract barcodes and product attributes from images and automated routine processing, reducing the department's product data processing time by about 40%.

• Developed low-bit quantization combining signed-Hadamard rotation with Output-Fisher recovery. Achieved INT2·INT3 PTQ SOTA: Llama-3 8B INT3 WikiText-2 6.99 (VPTQ 7.05), Qwen3.5-2B INT3 WikiText-2 11.09 and C4 15.62 (UniSVQ 12.62, 16.30), and Llama-3 8B INT2 at 2.265 bits WikiText-2 9.42.
• Implemented a llama.cpp C++/CUDA kernel that executes compressed weights directly, increasing decoding throughput 7.36× over GPTQ W3 (14.0 → 103.4 tok/s).

ICLR submitted

• Developed a method to improve WeCLIP+ under a research contract, combining CLIP and DINOv2 representations to repair disagreeing pixels and retrain with refined pseudo-labels.
• Trained the fusion head and decoder with frozen backbones, achieving WSSS SOTA in three months. Reached 53.31% COCO-Val mIoU, 1.5 percentage points above WeCLIP+.

• Built an MCP integrating web, code, and academic PDF search with full-text extraction. Combined exact, BM25, multilingual vector, and code/knowledge-graph retrieval through RRF and reranking to reuse collected material.
• Implemented source–chunk–evidence–claim relationships, ontology and data-lineage management, and Neo4j export. Delivered seven MCP tools for search and knowledge management, reaching the top 1% of MCP TOPLIST.

More Projects

• Monogram | Git-backed personal knowledge base built from shared links and documents, with hybrid retrieval and 13 MCP tools
• LangGraph Travel Agent | LangGraph travel agent delivered to a US travel agency with human review, HubSpot CRM, and Twilio follow-up
• E-AT | Skin-lesion classification with calibrated confidence; minimum ECE 1.03% vs CR-SAM (AAAI 2024) 1.7%, peak Macro-F1 0.8647