Time-Series Forecasting · LLM Agent Systems
Text should make a forecast explainable, not just better — and agents should finish the job.
↗ Visit my site — animated single-page portfolio (particle field, live counters, timeline).
Time-series forecasting. Why text helps a forecaster — and where it hurts. Multimodal modeling over numeric series plus repository / document text, with ablations that show which input actually moved the number.
LLM agent systems. Orchestrator + specialist workers, evidence-grounded reranking instead of free-form generation, RAG and tool-calling, shipped as containerized services.
Side track: cold-start cross-domain recommendation with multi-prototype uncertainty transfer.
Audit long documents for factual consistency: extract facts, catch numeric / logical / temporal contradictions, verify claims against the web, and check figures against text. FastAPI + React + Redis, Docker-ready.
parse → extract → detect → verify· MinHash pre-filter cuts pairwise comparisons 50–600× · one command to run.
A multi-agent travel planner: one sentence in, a mapped and budget-aware itinerary out. Orchestrator + specialist agents (culture, budget, route, guide, quality-eval) over AMap Web Services, rendered on Leaflet, with export to the TravelPlanner / TripTailor benchmarks.
Flask + Leaflet UI · agent state / memory / context kept per stage · benchmark export so plans are scored, not eyeballed.
An auditable session-analysis assistant for psychological counseling. Every observation is tied to evidence, severity comes from a rule table rather than a model, the LLM is reachable only at two named nodes, and clinician feedback is captured with six fixed reason codes.
7-stage pipeline (ingest → signals → chunks → events → agent → feedback → profile) · 23 hard rules · refuses cross-session claims below 0.7 comparability.
| Project | What it is | Result |
|---|---|---|
| GitPulse / OpenVista | 16 OpenDigger metrics × repo text, 32-month forecast + LLM explainer | MSE 0.0886 · R² 0.70 · DA 67.28% |
| factguardian | cloud-native fact-consistency agent for long docs | extract → cross-check → trace |
| Traveling-v2 | multi-agent itinerary planner over AMap + map UI | 7 agents · benchmark export |
| psycon-assist | auditable counseling session analysis | 23 rules · evidence-linked events |
| InteractiveFusion-Sentiment | image-text sentiment: dynamic-weighted IIF + sentiment-prototype SIE | Acc 84.7% · F1 84.0% · +3.5pt |
| Work | Role | Status | Key result |
|---|---|---|---|
| CrossBridge · cross-domain rec | co-first | submitted · ICDM 2026 | negative transfer → 1/3 |
| Time-series & text · attribution of multimodal gains | first | in progress | which text input actually moves the forecast |
| InteractiveFusion-Sentiment | solo | complete | 84.7% / 84.0% F1 |
| Deepfake audio · dual-branch Conformer | third | under review · ASC | lower EER on CFAD |
Python C++ SQL TypeScript · PyTorch Transformers Mamba/SSM Conformer · LangChain LangGraph RAG LoRA · FastAPI React Docker Redis
19 public repositories — dataset pipelines, models, services and evaluation harnesses, all readable end to end.
The three projects above have good first issue labels and issue templates; contributions are welcome.
