Projects

Ten systems, prototype to production

Generative AI and real-time NLP augmenting thousands of contact-center agents, credit models decisioning loans at large scale, and the earlier research and big-data work that got me there.

Live demo · you can try this one

Live Call Copilot

A public rebuild of the shape of the systems below, running live in your browser. Talk into your mic — or play a sample call — and watch a two-voice transcript, self-drafting notes, RAG-retrieved procedure docs, and a sub-second frustration alert. Deepgram + OpenAI composed into a real-time product; no proprietary anything.

Next.js Deepgram OpenAI pgvector Railway

Open the live demo Read the write-up

agentassistdemo.harrisonjansma.com

Live Call Copilot — two-voice transcript, self-drafting notes, RAG procedure docs, and a frustration alert

Generative AI & Real-Time NLP

Live, LLM-powered systems augmenting thousands of contact-center agents at Capital One.

01 Generative AI Real-Time Call Summarization GPT-4 Kafka Python AWS Lambda An agentic GPT reasoning pipeline on Kafka that auto-drafts servicing-agent notes mid-call, with an LLM "agent-as-a-judge" review stage before anything reaches the agent. Real-timeMid-call notes AgenticDraft + review In pilotGoverned rollout 02 Retrieval · RAG Live Procedure RAG RAG pgvector Embeddings Python A retrieval-augmented generation system that runs vector-embedding retrieval over live call transcripts to surface the right procedure and training documents to agents in real time. Real-timeMid-conversation Vector searchProcedure docs 03 Real-Time NLP LLM Complaint Detection LLMs Prompt Eval Python An LLM system that detects and categorizes customer complaints on live transcripts and flags escalations — lifting complaint capture sharply over a legacy frustration-phrase flag, at high precision. Earned an internal excellence award; pilot planned for 2026. GenAIComplaint capture High precisionvs legacy flag 04 Real-Time NLP Frustrated-Customer Alerting DistilBERT PyTorch Kafka DynamoDB Capital One's first real-time AI alert system — streaming DistilBERT and LLM sentiment over live calls at scale, alerting managers to de-escalate in real time. Real-timeLive alerting At scaleAcross the business StreamingReal-time pipeline

Applied ML & Credit Risk

Models that decision loans at large scale in auto lending.

05 Credit ML Auto-Loan Underwriting ML Suite Python XGBoost scikit-learn Snowflake Engineered and maintained the machine-learning suite behind Capital One's auto-loan underwriting, finding signal in credit characteristics tied to loan performance to decision loans at large scale. Production MLCredit risk RegulatedHigh-volume 06 Reinforcement Learning Credit-Policy Simulation Engine Reinforcement Learning Python Simulation Built a simulation engine that generates synthetic loan applications to power a reinforcement-learning task that optimizes credit policy — testing decisions before they ever touch a real applicant. RLPolicy optimization SyntheticApplication simulation

Earlier Work

Where I sharpened the fundamentals — research, big data, and NLP.

07 Anomaly Detection · Sprint Billing-Failure Anomaly Detection PySpark Hive Python Anomaly Detection Designed a novel model-training framework for anomaly detection on customer billing data, and built Hive and PySpark pipelines aggregating millions of records. The framework drove a 5x increase in the types of billing failure the system could detect — earning the only extension across an 80+ intern class. 5xDetectable failure types MillionsRecords aggregated 08 NLP Research · UPenn Population-Scale Emotion Classification Python scikit-learn NLTK Web Scraping Partnered with University of Pennsylvania sociologists on a long-term, population-level NLP study — building an emotion-classification model and designing Python scrapers over millions of public social-media profiles to measure population-wide shifts in emotional state following major political events. MillionsProfiles processed ResearchSociology study 09 Teaching · General Assembly Data Science Instruction Python pandas scikit-learn Jupyter Taught core data-science concepts on-site at Intuit through General Assembly — leading courses for software and data engineers during graduate study, translating dense statistical ideas into practical intuition they could apply on the job. IntuitOn-site cohort Core DSFor engineers 10 NLP · Writing Self-Taught Projects & Technical Writing LDA scikit-learn KMeans Medium Taught myself the field in the open — unsupervised LDA topic modelling that isolated toxic Wikipedia comments into their own cluster, KMeans/agglomerative clustering on survey data, and widely-read data-science writing on freeCodeCamp and Towards Data Science. 2 articlesfreeCodeCamp · TDS LDA · NLPTopic modelling