Work
Work from the lab
The projects are pieces of one operating direction: identity, tools, retrieval, evaluation, local models, software systems, and coordination. Public work is linked where possible; client work stays private.
Agent infrastructure
Identity, coordination, and connections to the outside world.
Agent identity infrastructure
TrustChain
Agent identity and trust infrastructure built around signed interaction records. The system brings together delegation, revocation, graph-based trust computation, an HTTP sidecar, and integrations with agent frameworks.
Implemented in Rust, with Python and TypeScript SDKs and cross-language verification tests. Public proof includes the TrustChain repository and draft-viftode-trustchain-trust-01, an individual Internet-Draft. It is work in progress, not a standard or endorsement.
Multi-agent systems
An operating layer for agent teams
A local orchestration system that turns a repository goal into routed tasks, agent prompts, work packets, and reviewable artifacts. Agents participate through adapters, with a record of assignments, work, and verification.
Implemented preview runs, portable work packets, adapter rehearsals, and offline verification. This is part of how the lab runs itself: Vlad directs the work; agents research, plan, build, review, and preserve evidence.
University workflow connectors
Personal MCP connectors
MCP tools that bring course material, deadlines, calendars, and documents into an agent conversation. The work includes account-bound sessions, silent refresh, local full-text search, source references, and previews before submissions.
Built as personal university connectors, with a connection overview, per-service diagnostics, account checks, source references, and explicit confirmation for changes.
Operating software
The applications and workflows around the models.
Private operations platform
Document, approval, and reporting flows
A full-stack platform for document workflows, risk registers, objectives, assessment campaigns, approvals, and reporting. Java and Spring on the backend; React and TypeScript on the frontend; PostgreSQL for the data.
Role-based access, database migrations, generated documents, audit trails, backend tests, and browser-level workflow tests. The software and its operating details stay private.
Knowledge assistant prototype
Source grounded retrieval
A study-assistant prototype that turns course PDFs into searchable material and answers with page references, figures, and confidence states. The pipeline combines OCR, document structure, keyword and vector retrieval, and checks on answer support.
The same work connects ingestion and retrieval to the product: streamed answers, citation cards, saved conversations, and feedback.
On-device machine learning
A local media lab
An Apple Silicon media workflow combining local models, a phone-upload interface, queued jobs, and side-by-side result selection. It connects model experimentation to a usable application, without sending media through a hosted inference service.
The engineering covers CoreML inference, retryable jobs, model comparison, and human review of outputs.
Research and evaluation
Experiments, reproductions, and systems put to the test.
Applied research
Diffusion and federated learning experiments
TU Delft course research reproducing diffusion sampling methods, including DDIM, Align Your Steps, and Zigzag Sampling. A separate federated-learning project investigates RoLoRA, alternating low-rank updates, and the training setup needed to reproduce a result.
The diffusion side covers sampling methods and image restoration experiments. The federated-learning side covers optimizer audits, ablations, and reproducible run records around RoLoRA adaptation.
Research project
Recommendation and ranking
Hybrid recommender experiments combining collaborative, content, and graph-based models. The work compares rank fusion, model switching, and contextual weighting, alongside diversity, calibration, coverage, and exposure.
Course research with frozen inference, matched baselines, held-out comparisons, and recorded negative results.
Software testing and reverse engineering
Learning how software breaks
TU Delft project work in fuzzing, symbolic execution, genetic program repair, and learning state-machine models from software behaviour. A shared benchmark compares random search, hill climbing, concolic execution, and AFL.
Seeded runs, matched time budgets, branch-coverage measurements, reachable-error checks, and a reproducible report pipeline.
AI Cup 2026
Bird movement under competition pressure
Vlad-George Iftode appears on the official Epoch AI Cup 2026 winners page as part of team Echo, fifth place, for the bird-movement AI challenge.
Presented as a team competition result, not a solo award or company credential. It fits the lab's research direction: model behaviour, evaluation, uncertainty, and fast technical judgment under a real benchmark.