Illia Dykanskyi

Architected, built, shipped, and operate a privacy-focused media analytics and redaction product end-to-end.

• Built the speech-to-text, policy matching, entity detection, and time-aligned redaction pipeline
• Engineered resilient high-throughput processing with automatic job recovery
• Built the complete responsive product experience from onboarding and upload through review, results, and billing
• Integrated subscription billing, credit balances, idempotent webhooks, refunds, and disputes
• Own provisioning, deployment, security hardening, monitoring, and ongoing production maintenance
• Designed privacy-preserving analytics that never capture user content
• Live at sayanara.com as a public audio alpha

Turned an unstructured concept into a production-minded restaurant recommendation platform and fundraising MVP in a three-person team.

• Partnered directly with the founder on roadmap, architecture, specifications, and core product decisions
• Worked full-cycle across frontend, data engineering, analytics, system design, implementation, and testing
• Normalized and deduplicated millions of records from 100+ sources into 200k+ enriched establishments
• Built multi-stage RAG ranking with vector search, reranking, preferences, and behavioral signals
• Designed a persistent-memory AI concierge with custom MCP tools
• Produced 170+ ADRs and implemented Prometheus/Grafana observability

Led five engineers building a modular KYC platform with a stateless ML core.

• Built document recognition, classification and extraction, liveness, face matching, deepfake detection, and tampering detection
• Reduced ongoing verification cost by approximately 100× by replacing proprietary vendor dependencies
• Took the system from concept to a production-ready MVP demonstration in one month
• Structured research, development, delivery, and long-term technical planning for a team of five
• Developed a Python SDK that reduced client integration time from weeks to days
• Presented technical roadmaps and delivery progress to C-level stakeholders

Operated as a cross-functional IC across data, ML, fraud analytics, internal tooling, and product operations.

• Built low-latency replication of 2.5+ TB from MySQL to BigQuery, reducing heavy queries from days or weeks to seconds
• Built an 800+ line anti-cheat analytics pipeline with ≈90% precision, zero false-positive bans, and automated enforcement for gross violators
• Created a centralized database API that removed inconsistencies across five-plus dashboards and tools
• Delivered a Telegram admin bot with real-time BigQuery analytics, role-based access, and controlled production mutations
• Built an NER/NLP invoice parser for chat payments
• Authored the first centralized documentation for 100+ tables, cutting analyst onboarding from weeks to days
• Supported game designers with economy balancing and currency-denomination mathematics

Built the data and AI layer for a crypto-market intelligence platform alongside a full-time role.

• Built Twitter and YouTube ingestion with NER-based semantic analysis
• Implemented a RAG research interface over collected market intelligence
• Created intent routing across specialized pipelines and a tool-using LLM agent for SQL, charting, and ML predictions
• Trained CatBoost, XGBoost, and AutoML forecasting models on Vertex AI

Built open-source multimodal dialog systems, audio tooling, model optimization, and agent integrations for the Dream ecosystem.

• Created dream_voice for ASR, TTS, and audio captioning and led dataset, fine-tuning, and benchmarking research
• Distilled CLIP to 50% of its original size with only 2–3% accuracy loss for faster production inference
• Extended dp-agent with Telegram video/audio handling and streaming responses
• Contributed services, annotators, and pipelines across dream_multimodal, dream_mint, dream_ocean, and dream_robot
• Built a ROS/Minecraft agent using LLaVa and Fromage with recursive goal decomposition and feedback
• Delivered 10+ merged OSS contributions and productionized modules from the master’s thesis

Built data tooling for a SaaS that turns product photos into marketplace-ready imagery, alongside a full-time role.

• Created a proprietary OpenPose C++ labeling pipeline that extracted pose data from unlabeled photos
• Managed Stable Diffusion fine-tuning experiments across hyperparameter configurations

Developed, benchmarked, and deployed production computer-vision systems for safety monitoring and industrial automation.

• Standardized detector and tracker evaluation across YOLO v3–v8, EfficientDet, and DeepSORT, cutting comparison time from days to hours
• Fine-tuned fire, smoke, PPE, and prohibited-symbol detectors and owned collection, cleaning, labeling, and tuning
• Built an open-source C# annotation tool with YOLO export that accelerated labeling by 3×
• Delivered an edge conveyor counter with <1% error at ≈30 FPS on 1080p video
• Load-tested multi-camera streams and profiled throughput and latency bottlenecks

Built backend features, admin tools, integrations, and performance improvements for production web systems.

• Implemented managed and composite caching for high-traffic pages, reducing TTFB by ≈25% and peak memory by ≈20%
• Built PHP/JavaScript backend features and administrative tooling
• Improved logging and alerting and resolved payment and form integration issues