Projects

Case studies from production AI systems.

Each entry leads with the outcome and what it enabled for the organization, followed by the architecture, security, and governance decisions that made it possible.

~/projects/hipaa-clinical-copilot

Case 01 · Behavior Frontiers · Regulated healthcare

HIPAA-compliant clinical agent platform

Outcome

Cut documentation turnaround for high-volume workflows, freed clinical time, and established an internal blueprint for safely extending LLMs to additional regulated workflows.

What it enabled

An agentic workflow platform that assists staff over PHI-safe knowledge bases, drafts intake and documentation artifacts, and routes anything ambiguous to a human reviewer.

Context

Clinical and operations teams spent hours navigating fragmented documentation, intake, and reporting workflows across PHI-heavy systems, with no safe path to apply LLMs.

Architecture

Retrieval layer over segmented, tenant-scoped indexes; tool-using agents orchestrated with typed contracts; evaluation harness with golden sets, red-team suites, and continuous regression tracking.

Security

PHI isolation, encryption in transit and at rest, least-privilege IAM, prompt-injection defenses, secret scanning, and full audit logging of model inputs, tool calls, and outputs.

Governance

Policy-driven guardrails, human-in-the-loop review, model and prompt versioning, evaluation gates before rollout, and structured incident response for model behavior regressions.

Technologies

PythonTypeScriptLangGraphOpenAIAnthropicPostgres / pgvectorAWSTerraform
~/projects/enterprise-rag-platform

Case 02 · Multi-tenant knowledge retrieval

Enterprise RAG platform

Outcome

Unified knowledge access across departments with measurable answer quality and no observed cross-tenant leakage in production.

What it enabled

A governed retrieval platform with per-tenant isolation, hybrid vector + keyword search, permission-aware chunking, and policy-checked answer synthesis.

Context

Teams across departments needed grounded answers over private document corpora, without leaking data across tenants or bypassing existing access controls.

Architecture

Ingestion pipeline with document normalization, chunking, and embedding; hybrid search over pgvector and OpenSearch; retrieval-time ACL enforcement; synthesis layer with citation-first prompting and refusal policies.

Security

Row-level tenant isolation, IAM-scoped access, encrypted embeddings, prompt-injection sanitization, and continuous evaluation against exfiltration probes.

Governance

Source citations on every answer, retrieval provenance logging, model routing policies, and evaluation dashboards for accuracy, refusal, and hallucination rates.

Technologies

pgvectorOpenSearchOpenAIBedrockTypeScriptTerraformAWS
~/projects/agentic-ops-copilots

Case 03 · Sales, RevOps & back-office automation

Agentic operations copilots

Outcome

Redirected significant operational hours to higher-leverage work with measurable throughput gains and no material incident record.

What it enabled

Tool-using agents that plan, act, and verify against enterprise systems, with structured evaluation harnesses ensuring behavior stays within policy as prompts and models evolve.

Context

High-volume, long-tail operational work (CRM hygiene, document drafting, inbox triage) consumed skilled time and resisted deterministic automation.

Architecture

Durable orchestration for long-running agent runs, typed tool interfaces to internal systems, retrieval for context, and offline evaluation against curated task suites.

Security

Scoped tool credentials, per-agent policy allowlists, action-level audit trails, and human approvals gating destructive operations.

Governance

Deterministic evaluation gates in CI, model version pinning, prompt change review, and post-deployment behavior monitoring.

Technologies

LangChainTemporalTypeScriptOpenAIAnthropicPostgres
~/projects/voice-ai-front-desk

Case 04 · Multi-location service operations

Voice AI front-desk & intake agent

Outcome

Recovered a large share of previously missed calls and shifted routine intake off human staff without measurable drop in caller satisfaction.

What it enabled

A production voice agent that answers calls, qualifies intent, books appointments, and escalates cleanly to a human — with structured call transcripts written back to the CRM.

Context

Front-desk staff were overwhelmed by inbound calls for scheduling, intake, and routine questions, causing missed calls and lost revenue outside business hours.

Architecture

Low-latency speech pipeline (STT → LLM planner → tool calls → TTS), telephony via SIP/Twilio, deterministic booking tools, and a supervisor model that scores every call for QA.

Security

Recording consent handling, PII redaction on transcripts, scoped API credentials per tenant, and rate-limited tool access to prevent misuse.

Governance

Per-call evaluation scoring, escalation logging, prompt and voice version pinning, and dashboards tracking answer rate, booking conversion, and containment.

Technologies

TwilioDeepgramElevenLabsOpenAI RealtimeNode.jsPostgres
~/projects/n8n-integrations-suite

Case 05 · SMB automation · Intelligent Integrations

AI-native integrations & workflow suite

Outcome

Replaced a fragile Zap sprawl with a governed automation layer, cutting integration incidents and enabling AI-assisted steps inside existing business processes.

What it enabled

A managed workflow layer built on n8n and typed adapters, augmented with LLM steps for classification, extraction, and drafting — with observability and replay built in.

Context

Growing teams were stitching CRMs, billing, email, and internal tools together with brittle Zaps that broke silently and had no visibility into failures.

Architecture

Self-hosted n8n on containerized infra, typed integration modules for Stripe / HubSpot / Gmail / Slack, LLM sub-workflows for enrichment and triage, and a run store for auditability.

Security

Per-workflow credential scoping, secret rotation, webhook signature verification, and structured error alerts to on-call.

Governance

Versioned workflows, staged rollouts, run-level audit trail, and SLOs tracked per integration for reliability and cost.

Technologies

n8nTypeScriptStripeHubSpotOpenAIDockerPostgres
~/projects/llm-evaluation-harness

Case 06 · Cross-team model quality tooling

LLM evaluation & regression harness

Outcome

Made model quality measurable and enforceable — teams ship prompt and model changes with confidence and a clear audit trail.

What it enabled

A reusable evaluation harness with golden datasets, rubric-graded judges, red-team suites, and CI gates that block regressions before they reach production.

Context

Prompt and model changes were shipping without a reliable way to catch regressions, and stakeholders had no shared view of model quality over time.

Architecture

Dataset registry with versioned test suites, deterministic + LLM-judge scoring, side-by-side model comparison, CI integration, and a dashboard for accuracy, cost, and latency trends.

Security

Scrubbed evaluation datasets, isolated evaluation credentials, and controlled access to sensitive golden sets.

Governance

Signed-off evaluation gates before promotion, historical scorecards per prompt/model version, and clear ownership of failing suites.

Technologies

PythonTypeScriptOpenAIAnthropicGitHub ActionsPostgres

More engineering work on GitHub

Architecture decisions, implementation details, and production-ready projects.

github.com/jbisaccia-9