Tap the right side for the next slide · tap the left side to go back · swipe or arrow keys work too

UNIVERSITY LECTURE & RESEARCH SEMINAR

Hermes Agent for Research Assistance

Autonomous Workflows, Literature Sweeps, Code Execution & Paper Drafting

Dr. Farshid Pirahansiah
AI & Computer Vision Engineer • pirahansiah.com

Why Hermes Agent for Researchers?

⚠️ Academic Research Pain Points

• 10,000+ papers/year per field (ArXiv overload).

• Manual literature synthesis & BibTeX management.

• Broken code repos, missing dependencies, failed builds.

• Passive LLM chats lose context and cannot touch local files/tools.

🚀 Hermes Agent Solution

• Autonomous tool calling: Terminal, ArXiv, Python, Web extraction.

• Persistent Memory & reusable Skill System.

• Parallel subagents (`delegate_task`) for concurrent sweeps.

• Runs on Local LLMs (oMLX/llama.cpp) & Cloud Providers.

A unified terminal & GUI agent that acts as a full-time research assistant.

Hermes Research Skills Catalog

8 new research skills installed + 8 existing — full toolkit for literature sweeps, critique & synthesis.

Skill	Trigger / what it does	Status
`deepdive`	Deep multi-source investigation; cited synthesis with confidence + disagreement mapping	NEW
`researchgap`	Finds gaps, contradictions & open problems in a topic's literature; proposes research questions	NEW
`critic`	Rigorous critique of a paper/claim/argument with severity-rated weaknesses + verdict	NEW
`literature-review`	Themed survey of a topic with annotated bibliography	NEW
`synthesise`	Merges several sources into one attributed synthesis; flags consensus vs open questions	NEW
`factcheck`	Verifies claims against sources; labels each supported/contradicted/unverifiable	NEW
`question-storm`	Generates diverse, prioritized research questions across levels	NEW
`cite`	Formats references in a citation style (APA/MLA/IEEE/BibTeX); finds real sources	NEW
`arxiv`	Search arXiv papers by keyword/author/category	existing
`llm-wiki`	Build/query interlinked markdown knowledge base	existing
`grounded-citations`	Ground answers in cited verifiable sources	existing
`reference-curation`	Audit/curate doc reference links	existing
`polymarket`	Query Polymarket markets/prices	existing
`competitor-news-monitor`	Watch companies for material news	existing
`blogwatcher`	Monitor blogs/RSS feeds	existing
`research-paper-writing`	Authoring structure for research papers	existing

Hermes Core Architecture

Skills

Procedural Memory
ArXiv, LaTeX, W&B, PyTorch, Git workflows.

Memory

Durable Profile
User bio, SSH nodes, GPU clusters, research domain.

Subagents

Parallel Workers
`delegate_task` for parallel experiments & sweeps.

Gateway

Multi-Surface
CLI, Desktop GUI, Web Dashboard, Telegram, Discord.

Supports 35+ providers: Local oMLX, vLLM, OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter.

Installation & Research Environment Setup

# 1. Official Hermes Shell Installer (uv + python venv + launcher)
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

# 2. Interactive setup wizard (Pick provider: Local LLM, Gemini, OpenAI, etc.)
hermes setup

# 3. Check environment health & dependencies
hermes doctor

# 4. Launch Hermes interactive CLI or Desktop GUI
hermes           # Terminal REPL
hermes desktop   # Native Electron GUI

Tailored Research Configuration

Configuring ~/.hermes/config.yaml for Research Labs

model:
  default: Qwen3.5-4B-OptiQ-4bit        # Fast local model for baseline task
  provider: custom
  base_url: "http://127.0.0.1:8000/v1"   # oMLX / llama-server local endpoint
  aliases:
    reasoning: gemini/gemini-2.0-flash-lite
    deep-coder: deepseek/deepseek-coder

agent:
  max_turns: 30
  context_compression: true            # Preserves state on huge paper reads

terminal:
  backend: local
  workdir: "/Users/farshid/research-projects"

Essential Hermes Skills for Academic Work

📚 Paper & Lit Review

• arxiv — Direct paper search & downloading.

• grounded-citations — Verified source grounding.

• llm-wiki — Karpathy-style Markdown Knowledge Base.

📊 Code & Execution

• jupyter-live-kernel — Live Python kernel.

• evaluating-llms-harness — Benchmark suites.

• systematic-debugging — 4-phase root cause analysis.

📝 Publishing & Docs

• docx / pdf — Parse & edit research papers.

• architecture-diagram — Dark-theme SVG diagrams.

• markdown-to-pdf — Export clean PDF preprints.

Workflow 1: Automated ArXiv Literature Sweeps

Command: "Find recent papers on 3D SLAM and build a literature matrix"

# Hermes executes ArXiv search -> parses PDFs -> builds references matrix
from hermes_tools import web_search, read_file, write_file

# 1. Search ArXiv by domain & topic
results = arxiv_search(query="cat:cs.CV AND title:SLAM", max_results=10)

# 2. Extract key contributions, equations, and benchmarks
matrix = []
for paper in results:
    pdf_text = web_extract(urls=[paper['pdf_url']])
    summary = extract_key_innovations(pdf_text)
    matrix.append({"title": paper['title'], "bibtex": paper['bib'], "notes": summary})

# 3. Output clean Markdown matrix + references.bib
write_file("literature_review.md", format_matrix(matrix))
write_file("references.bib", format_bibtex(matrix))

Workflow 2: Reproducing Code & Experiments

🔬 Automated Setup & Run

1. Clone GitHub Repo: Autonomous git checkout.

2. Inspect & Resolve Deps: Reads requirements.txt / environment.yml.

3. Fix Runtime Errors: Applies systematic-debugging to fix PyTorch CUDA mismatches or API changes.

4. Log Metrics: Generates W&B or TensorBoard plots.

💻 Live Tool Output

              $ hermes chat -q "Clone repository X, fix bugs, and run benchmark script"

              [tool call] terminal("git clone ...")

              [tool call] search_files(pattern="CUDA")

              [tool call] patch(file="model.py", old=..., new=...)

              [tool call] terminal("python eval.py --batch 32")

              ✓ Benchmark Complete: Accuracy = 94.2%

Workflow 3: Drafting LaTeX & Research Papers

1. Outline

Generates structured IEEE/ACM template with Abstract, Introduction, Methodology, Experiments, Related Work.

2. Citation Check

Uses grounded-citations to ensure every claim in the text links to a real ArXiv/DOI paper in references.bib.

3. Diagrams

Produces publication-ready SVG architecture diagrams, flowcharts, or Manim mathematical animations.

Zero hallucinated citations: claims are mechanically verified against PDF extractions.

Workflow 4: Parallel Research Swarms (`delegate_task`)

Spawning Concurrent Subagents for Heavy Academic Workloads

# Main Hermes Agent dispatches 3 subagents in parallel
delegate_task(tasks=[
  {
    "goal": "Search ArXiv for 2025-2026 Vision-Language-Action models and extract benchmarks",
    "context": "Focus on robotics manipulation"
  },
  {
    "goal": "Benchmark Qwen3.5 vs K2-Horizon on local M3 GPU using llama.cpp",
    "context": "Log tokens/sec, memory usage, and context latency"
  },
  {
    "goal": "Format collected findings into an IEEE LaTeX draft",
    "context": "Save to ~/paper_draft/main.tex"
  }
])

Persistent Memory & Custom Lab Skills

🧠 Persistent Memory (`memory`)

Stores facts that survive across every terminal session:

• Primary research focus (e.g. Computer Vision, 3D Reconstruction).

• Lab GPU Cluster details (e.g. SLURM partition a100-80gb, node IPs).

• Preferred paper format & writing style guidelines.

⚡ Custom Skills (`skill_manage`)

Teaches Hermes specialized lab procedures:

• Slurm Job Submission: sbatch script creation & queue monitoring.

• HuggingFace Dataset Uploads: Automated push of custom datasets.

• Lab Calibration Pipeline: Multi-camera camera intrinsic sweeps.

Cost & Token Efficiency Comparison

0 $

Local oMLX / Ollama
100% private, zero cost, runs on Mac/Linux.

90%

Prompt Caching
Free cache hits on long paper context blocks.

99%

Codebase Memory
Graph queries replace reading full repos.

10x

Speedup
Parallel subagent sweeps vs manual searching.

Local models handle draft tasks & code execution; Cloud models handle complex reasoning.

Case Study: Computer Vision & 3D Reconstruction

Project Goal

Build an end-to-end multi-camera 3D point cloud generation pipeline and publish open source benchmarks on pirahansiah.com.

Hermes Execution:

• Pulled 15 camera calibration papers from ArXiv.

• Debugged OpenCV / GStreamer C++ DMA pipeline.

• Generated benchmark tables and SVG pipeline diagrams.

Deliverables Produced

✓ 3d-vision.md documentation page on website.

✓ Verified OpenCV C++ code with zero memory leaks.

✓ BibTeX reference database with 15 verified citations.

✓ Interactive web visualizer demo.

Summary & Key Takeaways

✔ Autonomous Research Agent: Move from passive chatting to active tool execution.

✔ End-to-End Workflow: Literature review → Code reproduction → Experimentation → Paper drafting.

✔ Lab Privacy & Local Serving: Run on local oMLX / llama.cpp or Cloud APIs seamlessly.

✔ Parallel Power: Use subagents (`delegate_task`) for concurrent paper sweeps.

Explore the Knowledge Base & Code: pirahansiah.com

How to Write a Paper — Write in This Order

Draft bottom-up: build the evidence first, then frame it. Read top-down.

1 · Methods — Lock the protocol, setup, datasets & metrics. The foundation everything cites.
2 · Results — Report what happened, raw: numbers, curves, observed behaviour.
3 · Figures — Build the visuals: plots, diagrams, architecture art. One idea each.
4 · Tables — Assemble comparison/benchmark tables; numbers must match figures.
5 · Discussion — Interpret: why it works, limits, implications, relation to prior work.
6 · Introduction — Open with the gap & motivation; now you know what you proved.
7 · Abstract — Written last: the 200-word distillation of the finished paper.

Hermes: researchgap scopes the gap · critic stress-tests claims · cite formats refs · research-paper-writing structures the draft.

PaperBanana: Automating Academic Illustration

Text-to-figure AI for researchers — publication-ready diagrams & charts from a description.

🔧 What it does

• Methodology diagrams — model architectures, algorithm flows, encoders, system pipelines.

• Statistical plots — generates executable Matplotlib code from raw data (no numeric hallucination).

• Aesthetic enhancement — turns hand sketches / whiteboard notes into top-venue art.

• Styles — Transformer, GAN, RAG, Multi-Agent templates; posters & more.

🤖 How it works

• 5-agent closed loop (Planner → Visualizer → …) for faithful, precise, polished figures.

• Planner turns text into structured visual layouts.

• Visualizer renders via Nano-Banana-Pro — exact shapes, connectors, icons.

• Trusted by SNU, Stanford, UC Berkeley, CMU, Tsinghua, SJTU, Yonsei, IIT Madras, Monash, Broad Institute.

Example: prompt "encoder–decoder segmentation pipeline with skip connections" → ready architecture diagram in seconds. Reference: paper-banana.org

Academic Peer-Review Loop

From submission to camera-ready — and how Hermes shortens each cycle.

📤 Submit

• Format to venue template (IEEE/ACM).

• grounded-citations checks every claim links to a real source.

• cite exports clean BibTeX/references.

🔍 Review

• Editors assign 2–3 anonymous reviewers.

• critic pre-empts weaknesses before submission.

• factcheck flags unsupported claims.

✏️ Revise

• Address each point (accept / rebut).

• deepdive gathers new evidence for gaps.

• synthesise merges reviewer feedback.

Loop repeats per round (R1 → R2 → …) until accept; then camera-ready + research-paper-writing polish. Average: 2–4 months per round.

Agent Swarms: Parallel Research at Scale

One orchestrator fans out to many specialized workers, then merges results.

🐝 How a swarm runs

• Orchestrator splits the task into independent sub-goals.

• Swarm = N parallel subagents, each with its own context & tools.

• Merge: results aggregated, deduplicated, cross-checked.

• Hermes: delegate_task spawns concurrent workers (no shared context bloat).

🌊 Example: kimi.com → swarm

1. kimi.com receives the research question (e.g. "compare 2025 VLM agents").

2. → swarm: fans out to subagents — ArXiv sweep, benchmark mining, repo cloning.

3. → …: each worker returns findings; orchestrator synthesizes one report + BibTeX.

• Same pattern in Hermes: 3 delegate_task calls → one literature matrix.

Why: 10× speedup vs serial search, isolated failures, no single context overflow. Pair with synthesise + critic.

LeapSpace & Your Research Profile

Elsevier's research-grade AI workspace (Scopus data) — and the author record it surfaces.

🪐 LeapSpace (Elsevier)

• AI workspace powered by Scopus + Elsevier full-text.

• Deep Research reports: patterns, contradictions, evidence gaps.

• Trust Cards + Claim Radar: sources & contradictions shown.

• Writing Coach, Funding Scout, Author Search, collaborators.

• researcher.elsevier.com/eur

👤 Farshid Pirahansiah — Scopus

• h-index 6 · 12 publications · 73 citations

• Affiliation: Center for AI Technology, Bangi, Malaysia.

• Active years: 2010–2022.

• Topics: camera calibration, robot localization, image segmentation, OCR, license-plate recognition.

• Profile generated by LeapSpace, Sun Sep 06 2026.

Hermes mapping: deepdive/literature-review ≈ Deep Research · grounded-citations/factcheck ≈ Trust Cards/Claim Radar · cite ≈ Author Search/BibTeX.

Best Reference Finder — AI Citation Tools

Dedicated finders query Crossref / Semantic Scholar — no hallucinated refs like chatbots.

Tool	Best for	Standout	Styles
Sourcely	Bulk / paragraph	Scans essays, highlights text needing support, summaries	700+ (BibTeX, RIS)
Citely AI	Claim verification	Paste claim → matched to verified DOIs	APA, MLA
Grammarly	In-line writing	Flags missing sources in editor, inserts citations	APA, MLA, Chicago
Paperpal	MS Word flow	250M+ articles, search by title/DOI	10,000+
SciSpace	Partial metadata	Broken URL/author → complete citation	BibTeX, RIS, CSV
Writeless	PDF uploads	Filter by min year, inline insertion	APA, MLA, Harvard…

Beyond find: Scite labels citations support/challenge/mention · Google Scholar (free, manual) · Consensus (Consensus Meter) · manage with Zotero / Mendeley / MyBib. Hermes: cite formats, factcheck verifies, grounded-citations grounds.

Second Brain: Obsidian × Hermes

Turn your notes into a queryable, agent-editable knowledge base.

🧠 Connect Obsidian

• Vault = plain markdown; Hermes reads/writes via the obsidian skill.

• Path from OBSIDIAN_VAULT_PATH or ~/Documents/Obsidian Vault.

• Use [[wikilinks]] + YAML frontmatter (tags, aliases) for structure.

• Two-way sync: edit in Obsidian, Hermes sees it next turn.

🔌 Hermes skills for your vault

• obsidian — search, read, create, edit notes & wikilinks.

• llm-wiki — build interlinked markdown KB.

• synthesise — merge notes into one.

• cite — attach BibTeX to literature notes.

• factcheck / grounded-citations — verify before saving.

Flow: capture → Hermes tags & links → query with search_files → surface in deepdive / literature-review. Your second brain, always current.

Google Opal Experiment

Google's experimental AI app platform — verified facts only; function not yet documented publicly.

🔎 What is confirmed

• Official Google product labelled "Opal [Experiment]".

• Web app (opal.google) loading in a sandboxed iframe.

• Requires Google sign-in (OAuth) to use.

• Requests camera & microphone permissions.

❓ Not yet public

• Exact purpose / feature set is not in the page source.

• Positioned in Google's "experiments" track (like Labs).

• Likely an AI app-/agent-building surface — unconfirmed.

• Check opal.google after sign-in for live capability.

Hermes angle: if Opal is an app/agent builder, it parallels skill_manage (custom skills) & delegate_task (subagents) — but treat as speculation until docs ship.

Validated Skills Library installed & scanned

All installed via hermes skills install, security-scanned, and verified (valid SKILL.md, enabled).

Web & Research

agent-reach · youtube-full · defuddle · resemble-detect

Build & Engineering

using-agent-skills · setup-matt-pocock-skills · make-interfaces-feel-better · humanizer

Agents & Ops

browser-harness · i-have-adhd · loopy · loop-library · skillclaw

# list active skills hermes skills list # 14 validated community/url skills now active

Command Cheat Sheet DESK · CLI · MSG · SHELL

Everyday commands — Desktop, CLI chat, messaging, shell. (Verified vs current Hermes source, Sep 2026.)

Sessions & Context

/new · /resume · /sessions · /title · /branch · /compress · /context · /status

Control Work

/queue · /steer · /bg · /btw · /agents · /stop

Goals, Loops & Plans

/goal · /subgoal · /heartbeat · /loop · /plan · /review · /refine

Models & Behavior

/model · /moa · /personality · /reasoning · /fast · /approvals · /yolo · /busy · /voice

Skills, Memory & Tools

/skills · /learn · /memory · /init · /tools · /browser

Automation & Coord

/cron · /suggestions · /blueprint · /kanban

Inspect, Recover & Fix

/retry · /undo · /save · /diff · /rollback · /usage · /debug · /help

Messaging / Gateway

/sethome · /topic · /commands · /approve · /deny · /pause · /platform · /restart

Terminal Essentials

hermes / hermes chat · -z "prompt" · model · status · doctor · gateway status · --safe-mode · update

# repeat-work modes /goal = work until objective met /loop = repeat w/ stop conditions /heartbeat = one recurring check /cron = durable schedule outside chat # type / + letters for autocomplete; /help for full list

Thank You & Discussion