Selected work

Things I've built and shipped.

Production-grade AI systems — agentic pipelines and multi-agent applications, taken from notebook to deployment.

01 / AGENTIC PIPELINE

SmartIngest

An agentic pipeline that turns messy, unstructured documents into validated, routed structured data.

Python 3.12LangGraphFastAPIGeminiPydantic
72 testseval-gated CIDockerized
retry on low confidenceGuardrailsClassifyExtractValidateRoute

02 / MULTI-AGENT SYSTEM

RecruitRadar

A CrewAI multi-agent screener that scores resumes against a JD with deterministic, evidence-backed results.

PythonCrewAIGeminiLiteLLMrapidfuzz
~6 hrs saved / role0–100 deterministic scoreevidence-backed
JD AnalystResume ScorerInterviewDesignerOutreachDrafterTrackerEvidence guardrapidfuzz · every claim cites the resume

03 / CV SERVING PIPELINE

football-detect-serve

A football object detector (ball, goalkeeper, player, referee) trained on broadcast footage, exported to ONNX, quantized to INT8, and served behind a batching FastAPI service — every optimization measured through one evaluator, not assumed.

PythonUltralytics YOLO26PyTorchONNX RuntimeFastAPI
ONNX-fp32: +1.5x throughput, freeINT8: no CPU speedup, -0.03–0.06 mAPload-tested @ concurrency 1–16
TrainEvaluate (gate)Export ONNXQuantize INT8Serve (batch)Benchmark+ load testtorch vs onnx-fp32 vs onnx-int8, same evaluator

04 / SPORTS ANALYTICS

Football Tactical Metrics

Turns raw player-tracking data from two different providers into validated tactical and physical metrics, checked against synthetic ground truth.

Python 3.13kloppypandasSciPyShapely
168/190 checks pass22/22 synthetic checks95% coverage rule
Metrica /SkillCornerUnified schemaMetrics layerValidationDashboardCoverage rule≥95% tracked, or the value stays empty

More work

More projects on GitHub