Dani Szwarc
Montreal · AI Systems Engineer
I've spent two decades building enterprise software in environments where reliability isn't optional. Over the last few years I shifted focus to AI systems, not because it's trendy, but because that's where the hard problems are right now.
Most of my recent work sits at the intersection of document intelligence, RAG pipelines, and agentic automation. A recurring theme: AI systems that behave predictably under real operating conditions.
Close to the problem. I stay close to the user through the entire delivery. The system has to fit how people actually work, and that only happens through iteration.
Reliability over features. Audit logs, RBAC, prompt injection defense, anonymization pipelines. In the environments I work in, these are part of the architecture from day one.
Production first. Every project I build runs on real infrastructure. Docker, Traefik, CI/CD, self-managed VPS. Deployed and maintained.
Iterative with real users. Every project in this portfolio evolved through direct feedback cycles with the people who use it. That is just how good software gets built.
Dissertation project converting natural language descriptions into formally verified executable workflows using a two-layer DSL and LLM-guided iterative refinement. LangGraph state machine with Lark-based grammar, 25 constrained primitives, and a verify-then-repair loop targeting n8n and LangChain output.
82/82 unit tests · formal verification · University of Liverpool
View on GitHub →End-to-end automation system for an Argentine insurance brokerage with 5 producers. WhatsApp messages (text and audio) and emails are automatically ingested, transcribed (Whisper), classified by AI, and surfaced as tasks on a shared Kanban dashboard. Producers manage policy activations and cancellations without manual triage.
WhatsApp + email intake · AI classification · Kanban dashboard · TOTP 2FA
SOX-compliant internal knowledge platform for a Fortune 500 client. RAG-powered chat interface with on-premises LLM inference, Markdown-to-HTML pipeline, full audit log UI, and RBAC. Worked directly with internal stakeholders through iterative delivery cycles.
SOX-compliant · Fortune 500 · on-premises inference
Production clinical AI platform processing 289+ anonymized medical documents into a unified intelligence dashboard with RAG-powered assistant, mandatory source citations, and 4-layer prompt injection defense.
RBAC · audit log · on-prem VPS · bilingual EN/ES
Multilingual AI email agent for a Montreal community soccer club. Automates responses to player and parent inquiries in EN/FR/ES using RAG over club documents. OAuth-authenticated Gmail integration deployed in production.
Production · trilingual · zero-downtime model swap
AI pipeline automating 40+ articles per issue from PDF to Drupal CMS for an economics cooperative. Claude Vision API + n8n orchestration; taxonomy matching with accent normalization. Delivered iteratively with the client.
30 min → under 3 min per article · 10x+ efficiency gain
Full-stack mobile fitness tracking app with AI food label scanning (vision model), workout logging, and progress stats. Built end to end as a personal product.
Personal product · end-to-end ownership · AI vision integration
Platform ingesting patient medical histories (PDFs) into structured clinical knowledge wikis using adversarial dual-model iteration (Claude Opus + GPT-4o) until clinical findings converge. Taken private; shared as architectural case study.
80% LLM cost reduction · adversarial dual-model · compliance-by-design
Installable offline-capable PWA that extracts and reads French academic PDFs aloud. pdf.js handles client-side text extraction with heading detection based on font-size ratios. A FastAPI microservice on a self-managed VPS proxies extracted text to an LLM for semantic cleanup before playback via the browser's native Web Speech API, with real-time sentence-by-sentence highlighting.
PWA installable · offline-capable · GitHub Pages + self-managed VPS
View project →Full-stack delivery, AI solutions architecture · Montreal
End-to-end product delivery for enterprise clients across healthcare, financial services, and manufacturing. Built a mobile-first PWA for on-site nurses (ColdFusion/REST/MSSQL), a medical rehabilitation billing system (60% reduction in manual processing), and an enterprise ticketing system (40% reduction in cross-department resolution time).
Fresche Solutions · Montreal
Lead modernization of legacy systems for enterprise clients. Technical bridge between legacy business logic and modern stack — translating complex 4GL/COBOL workflows into PHP/JavaScript/SQL architectures.
University of Liverpool
Dissertation: WorkflowSynth — LLM-guided programme synthesis with formal verification for enterprise workflows
Centennial College, Toronto
Certifications
AI / LLM
LangChain · LangGraph · RAG · pgvector · Agentic workflows · OpenAI · Anthropic · n8n
Production
FastAPI · PostgreSQL · Docker · Traefik · CI/CD · GitHub Actions
Frontend
React · Next.js · TypeScript · JavaScript · React Native · Expo
Languages
English (fluent) · Spanish (fluent) · French (basic)