Projects

Selected work across AI platforms, agents, and production systems. Read the case studies.

AI Sales Agent (Purchase Advisor - Edtech)

AI Builder · Machine Learning · AI Bot · RAG · Hybrid Retrieval · ETL

What I delivered

Built RUBI, an embedded RAG purchase advisor on OnCourseLearning.com. Customers get fast, cited answers on courses, licensing, and state-specific policies from a large, changing catalog. Delivered ETL + multimodal indexing, hybrid retrieval (metadata + vector + structured lookups), grounded answers, purchase flows, analytics, and tuning loops — so support isn’t stuck on repetitive questions.

Impact

695,042 conversations served; ~2.6 questions per convo; ~10s to first token, ~11s to complete answer. Knowledge base >100k structured rows. Feedback: 6% up / 94% down.

Users

700K end-customer conversations on OnCourseLearning.com

Revenue generated

$900K influenced by RUBI-assisted purchases and support cost avoidance

AI powered Data Labeling Agent

AI Agent Development · Artificial Intelligence · AI Bot · Data Labeling · Machine Learning

What I delivered

Security teams lacked fast, consistent labeling of diverse telemetry; manual triage + spreadsheets slowed model improvement and left attack-coverage gaps. Built a labeling pipeline that ingests/normalizes telemetry, retrieves playbook context, uses GPT-4.1/5 to draft Pydantic-validated labels, and applies QA + consensus checks with human review only for low-confidence cases

Impact

Thousands of structured labels; ~70% reduction in human review; retrain cycles cut from weeks to days; contributed to major incident-reduction gains (up to 97% per client metrics)

SLM-PORTAL— Finetuning Platform

What I delivered

All-in-one platform that turns your raw documents into specialized, fine-tuned small language models without needing an ML team. You simply upload your files, auto-generate clean training datasets, fine-tune a lightweight model, and deploy it securely via OpenAI-compatible API. It keeps your organization's data private with complete tenant isolation while remaining significantly faster and cheaper than running giant LLMs.

Impact

Empowers domain experts to build, fine-tune, and deploy private custom AI agents from documents without an ML team. Cuts operational costs and complexity by replacing massive models with fast, secure, and fully owned small language models.

Project link

Agentic Trading Platform

AI Trading · Trading Automation · AI Agent Development · Claude · Machine Learning

What I delivered

An always-on, agent-native autonomous trading system integrating a brokerage-execution MCP. Dual-loop design — a fast “tick” loop that reads live account/market state and only trades through a mandatory review→place-order pipeline, plus a separate slow loop that backtests and evolves strategy.

Impact

98% win rate · 100% uptime

Users

389 accounts

Revenue generated

$2599

Project link

Agentic RAG - Permit Signal

What I delivered

Ingested and normalized 3M NJ construction permits and 60K properties from Socrata and ArcGIS, geocoded into PostGIS, and built a hybrid retriever (spatial → semantic → BM25/SQL) behind a chat-style “Signal Agent” that returns cited answers, map pins, and exportable results.

Impact

Turned fragmented, slow permit/property lookups into instant, source-linked conversational search across an entire state’s dataset.

AI Quoting Assistant (AOI Marketing RFP Response Generator)

AI · RAG · chatbots

What I delivered

OCR ingestion of internal PDFs/images, content normalization and chunking, a retrieval layer, and a RAG pipeline that generates RFP responses grounded only in approved internal materials.

Impact

Cut RFP first-draft turnaround from ~5 business days to under 4 hours, and reduced review cycles from 3–4 rounds to usually 1.

Users

20+ person sales and proposal team at AOI Marketing (account managers + technical writers).

Revenue generated

Supported ~$2.4M in RFP pipeline in the first two quarters; contributed to 3 closed deals totaling ~$680K where the tool drafted the initial response.

AI Security Platform — Prompt Injection Defense

Python · FastAPI · Envoy · Pydantic · GPT-4.1 · classifiers · Redis

What I delivered

A protective sidecar gateway in front of LLM endpoints — multi-layer detection (regex/AST, ML classifier, semantic checks), a GPT-4.1 sanitizer, and a deterministic Pydantic policy engine.

Impact

>95% attack detection, <3% false positives, <40ms median added latency, ~87% fewer AI-content incidents.

Users

200+ protected endpoints/day

Revenue generated

800K-1M

Fine Tuning - Cybersecurity LLM

What I delivered

Built a lightweight, high-performance security labeling model that detects and categorizes adversarial LLM inputs. Fine-tuned Gemma 7B with QLoRA to flag prompt-injection and SQL-injection attempts with high precision while keeping a minimal hardware footprint.

Impact

96% accuracy , precision 96% , recall 98%

Users

teams using the classifier for protected endpoints

RFP Generator

What I delivered

Built an agentic RFP Response Generator that answers RFP requirements about the company’s product family using only verified internal materials. OCR extracts text from PDFs/images (datasheets, install guides, technical guides, brand materials, prior RFP examples), then content is normalized, chunked, and indexed. A RAG pipeline retrieves the most relevant approved passages per requirement and generates a concise, grounded response.

Impact

Cut average RFP draft time from ~3 days to under 4 hours; ~80% of requirements auto-drafted from approved internal sources with cited passages.

Users

Sales and proposal team (8–12 users) drafting responses from verified product docs

Revenue generated

Supported proposals tied to multi-six-figure deal pipeline; faster grounded drafts improved response coverage and bid readiness.

OpenClaw - AI Assistant

Machine Learning · AI Chatbot · AI Bot · AI App Development · AI Agent Development

What I delivered

Built OpenClaw, a smart task manager where AI agents analyze and prioritize work. Tasks go through AI review for prioritization, break down into subtasks from each card, and get suggested next actions — then OpenClaw recommends what the agents should work on next.

Impact

time saved per planning cycle, 90% of tasks auto-prioritized, or fewer missed deadlines

Users

internal teams

Revenue generated

saves the $450/month

Data Analysis Agent

Data Analysis Agents · Copilot · Automated Reporting

What I delivered

Built and deployed enterprise data analysis agents (including Copilot-based workflows) for ops teams. Agents ingest operational data, run autonomous cleaning, surface analytics, and generate automated weekly reports and alerts — turning scattered ops data into centralized, actionable insights across departments.

Impact

Cut weekly reporting from ~8–12 hours of manual work to an automated run; ops teams get consistent insights and alerts without waiting on ad-hoc analysis.

Users

Multiple operations teams across departments using a shared automated analytics / weekly-report pipeline.

Revenue generated

Freed analyst capacity and faster decisions; estimated mid–five-figure annual savings from reduced manual reporting and earlier ops issue detection.

AI Report Generation Agent

Python · Tesseract · Slack API · OpenAI GPT · Celery · S3

What I delivered

Marketing teams couldn’t quickly find or reuse product content scattered across PDFs and briefs. Solution: Built a PDF-to-Slack RAG pipeline: auto-ingest (Drive/Dropbox/S3), layout-aware OCR + extraction, NER normalization, vector DB + Postgres indexing, and Slack slash commands to fetch snippets and generate briefs/ad copy.

Impact

Centralized KB, much faster campaign prep, better reuse/consistency.

AI Revenue Reporting Agent - EdTech

AI Agent Development · AI Development · Artificial Intelligence · Agent GPT · AI Model Training

What I delivered

Revenue data was fragmented across LMS, payment, and accounting systems; manual reconciliation of refunds/prorations made reporting slow and error-prone. Solution: Built a Revenue Reporting Agent to ingest and normalize feeds, apply AI-assisted mapping, validate/reconcile transactions, flag anomalies, and push curated datasets to Power BI.

Impact

Faster daily/weekly closes, up to ~90% less manual reporting effort, better anomaly detection and auditability.

AI Document Processing Agent - HealthTech

AI Agent Development · AI Development · Document Processing Software · Artificial Intelligence · AI-Enhanced Classification

What I delivered

Mixed-quality PDFs/scans/notes required manual cleaning + extraction for downstream workflows, creating slow, error-prone, recurring ops cost. Deployed an AI doc-processing agent to run OCR, layout-aware extraction, entity normalization, and cleaning; added Pydantic validation plus human-in-the-loop review for low-confidence fields; integrated structured outputs into reporting/analytics systems.

Impact

Cut manual processing time dramatically $1M annual revenue + cost savings across deployments; high accuracy on common fields with review focused on edge cases

Users

healthcare services providers