Dumitru Luca | Portfolio & AI Interactive Profile

YouTubeTranscriptGen.com LIVE

Production web app providing full, multi‑language YouTube transcripts with a streaming AI assistant for question answering & summarization. Emphasis on resilience and UX speed.

Transcript extraction + language switching (cached, fallback tolerant)
Streaming AI chat (incremental markdown, scroll anchoring)
Exports: TXT / SRT / VTT / JSON; ratings + credit economy
Proxy rotation & warm path logic to mitigate rate / geo constraints
Stripe billing + structured data SEO + persisted conversations

Tech focus: Efficient transcript retrieval pipeline, resilient network layer with rotating proxies, incremental SSE rendering for near real‑time UX, token cost control strategies.

Challenges solved: Handling regional blocking, balancing model latency vs quality, preventing scroll jank during long streaming outputs.

Impact: Enables users to extract and query long-form video knowledge in seconds instead of manually scrubbing hours of content.

Visit Project ↗

Just for You (Java 17)

Terminal-based messaging & daily interaction app built in Java 17. Private two-user ecosystem spanning scheduled notes, real-time chat, and lightweight games over a self-hosted socket server.

Main menu launcher stitches modular sub-apps (Messages, Daily Note, Games, etc.) inside a single terminal session.
Hetzner-hosted Java socket server manages message exchange, chronological ordering, and daily delivery windows.
Custom timestamped protocol (timestamp username message) keeps ordering deterministic and sortable server-side.
File-backed persistence + server warm-start logic preserves history and scheduling metadata across restarts.

Highlights: Real-time messaging between two remote clients, daily message drops, and shared mini-apps that run without leaving the terminal.

Reliability: Pure Java I/O stack with minimal dependencies, layered atop file persistence so the server can pre-load state and recover gracefully.

Impact: A fully self-hosted, private communication environment engineered for simplicity, emotional connection, and day-to-day reliability.

Retrieval‑Augmented QA Prototype

Experimental pipeline for answering domain questions over large text corpora (e.g. legal). Focus on retrieval relevance & latency budget control.

Document chunking + embedding & vector store indexing
Hybrid similarity retrieval + context window packing heuristics
Deterministic assembly layer feeding generation model
Instrumentation for recall / precision & token efficiency

Approach: Hybrid scoring (semantic + term features), rank truncation, adaptive window packing based on token budget.

Evaluation: Early harness for manual relevance judgments + recall tracking across query set.

Next: Add multi‑vector representation and shallow re‑rank pass.

Router Data Plane Simulator (C)

Low‑level packet processing path built over Mininet to internalize protocol behavior & failure modes.

Ethernet frame parsing, ARP cache + pending queue
IPv4 TTL decrement & checksum recomputation
Longest Prefix Match routing decisions
ICMP: unreachable, time exceeded, echo reply logic

Focus: Reinforce deterministic packet path reasoning, timing of ARP resolution vs queued packets, and correctness of checksum updates.

Learning: Deepened understanding of failure surfaces (ARP exhaustion, TTL expirations) & diagnostics flows (ICMP signalling).

Outcome: Better intuition for latency sources and error propagation across network layers.

Visit Project ↗

I'm Luca.

Know About Me

Hobbies & Interests

Skills

Projects

YouTubeTranscriptGen.com LIVE

Just for You (Java 17)

Retrieval‑Augmented QA Prototype

Router Data Plane Simulator (C)

AI Assistant About Me

Contact

Email

LinkedIn

Resume

Phone