Engineering Blog

Deep-dives from our engineers — real production builds, the architecture behind them, and the measured impact.

🇵🇱Igor · Full-Stack & SaaS Architect · presenting for Oddly Automated

How Oddly Automated ships production AI systems

Same team, same standard — production AI systems spanning autonomous due diligence, multi-tenant SaaS, and on-premise AI copilots. The three projects below are the ones featured in the showreel.

Read the full transcript

Hi, I’m Igor, from Oddly Automated — an EU-based AI engineering team that builds production systems which actually ship. For corporate legal and finance teams, we automated due diligence end to end: our OSINT pipeline on n8n, Python, and Claude audits over 10 company registries in under 25 seconds and returns a sandboxed risk report. For digital agencies, we engineered Chill CRM — a multi-tenant PSA platform with strict schema isolation and automated Stripe billing that lifted client margins by 28%. And for education, we built prismOS — an adaptive teacher copilot that pairs Gemini 2.5 Pro with a self-hosted local model for full data privacy, saving teachers up to 70% of their prep time. Same team, same standard: clean architecture, real tests, measured impact. Explore the projects below.

Projects featured in the showreel

AI OSINTEnterprise Risk ManagementRegulatory Automation

Autonomous OSINT Due Diligence — Counterparty & Company Audit

🇮🇪 Denis · Automation & OSINT Engineer · Ireland

The Challenge

Corporate legal and procurement departments waste dozens of manual hours cross-referencing fragmented government registers, court databases, and corporate ownership filings. Manual verification creates severe operational bottlenecks and exposes organizations to shell companies, undisclosed liabilities, and toxic counterparties.

Solution & Architecture
  1. Parallel Ingestion: Automated scrapers query 10+ public and statutory company registries concurrently.
  2. Entity Resolution: Python scripts normalize registry data, track ownership trees, and reconcile discrepancies.
  3. Semantic Risk Scoring: LLM evaluation (Claude 3.5 Sonnet) analyzes filings for red flags, political exposure, and insolvency signals.
  4. Sandboxed Delivery: Dynamic generation of an interactive HTML report rendered in a sandboxed iframe.
Key Metrics & Impact
< 25s
audit latency (from 2-3 hours)
10+
registries queried
100%
sandboxed execution
Tech Stack
  • n8n
  • Python 3.11
  • OpenRouter API
  • Claude 3.5 Sonnet
  • Tailwind CSS
  • Sandboxed Iframe
Enterprise B2B SaaSProfessional Services Automation (PSA)

Chill CRM — Multi-Tenant Agency PSA & Profitability Platform

🇵🇱 Igor · Full-Stack & SaaS Architect · Poland

Open live product →
The Challenge

Digital agencies routinely lose up to 25% of their billable revenue. Disjointed timers, untracked out-of-scope tasks, and delayed invoicing erode project margins and burden project managers with manual administrative tasks.

Solution & Architecture
  1. Multi-Tenant Schema Isolation: Dedicated PostgreSQL schema per tenant, guaranteeing zero cross-customer data leakage and strict GDPR compliance.
  2. Automated Financial Operations: Real-time retainer burn forecasting, hierarchical rate cards, and direct Stripe Billing integration.
  3. External Spend Synchronization: Live bi-directional integration with major ad-platform APIs (Meta, Google Ads).
  4. Global Readiness: Native multi-language localization (EN, UK, RU, PL).
Key Metrics & Impact
+28%
agency margin increase
15+ hrs/week
saved on PM
EU / US
live in production
Tech Stack
  • Next.js 15
  • TypeScript
  • Node.js REST
  • PostgreSQL (Schema Isolation)
  • Stripe Billing
  • Ad-Platform APIs
EdTech AIEnterprise Agent OrchestrationClean Architecture

prismOS — Adaptive AI Teacher Copilot with On-Premise Isolation

🇩🇪 Valera · AI Systems & Backend Engineer · Germany

The Challenge

Teachers spend up to 65% of their work week on administrative overhead: drafting bespoke test materials, manual grading, and tailoring lesson plans. Using consumer AI raises severe data-privacy violations and exposes schools to unvalidated hallucinations.

Solution & Architecture
  1. Hybrid Multi-Model Orchestration: Intelligent routing between Gemini 2.5 Pro (complex reasoning) and self-hosted local Qwen 2.5 (sensitive student PII).
  2. Strict Data Privacy: Complete isolation of student PII, meeting German DSGVO / GDPR requirements.
  3. Low-Latency Streaming: Bi-directional WebSocket channels for instant token streaming.
  4. Test-Driven Reliability: 157 pytest integration test cases enforcing schema validation via Pydantic.
Key Metrics & Impact
up to 70%
teacher hours saved
157
automated pytest cases
100%
DSGVO-compliant local isolation
Tech Stack
  • Python 3.12
  • FastAPI
  • Pydantic V2
  • Gemini 2.5 Pro
  • Local Qwen 2.5
  • WebSockets
  • Docker
  • Pytest