

Building Agentic Systems
& Full-Stack AI Software.
I'm William Volodarsky, a Software & AI Engineer (BS CS, Fordham, Magna Cum Laude). I build production web apps, agentic workflows, and internal tooling in Python, TypeScript and SQL, backed by 8+ years of real-world operations and clinical data work.
pytest unit + FastAPI integration tests (HITL commits, branching, engine)
129M of 156M input tokens served from cache across 476 logged runs
Gmail, Calendar, web search, Chrome bridge, HL7 parser, ADB scraper and more
Markdown vault navigated with filesystem tools; no vector index
Nexus: An Agent-Integrated Knowledge OS
A full-stack LangGraph + FastAPI + Next.js system I built and run for myself to query, maintain, and automate a ~400-note Markdown knowledge base, plus my email, calendar, and documents. The engine is public; the personal vault content stays local.
Filesystem navigation instead of a vector index
Behavior lives in Master Instructions.md → Section Instructions.md → per-note YAML, editable in Obsidian and live on the next turn.
- Grounded paths. The model sees the real folder tree and tool catalog up front, then reads whole notes with read_file and search_vault rather than retrieving 500-token chunks.
- No re-indexing. Notes I edit by hand are visible on the very next call; there is no embedding pipeline to fall out of sync.
- Cache-friendly by design. The system prompt is byte-stable within the hour and history is never rewritten, which is why 82.8% of input tokens are served from the provider's prompt cache.
- Tradeoff: whole-note reads cost more tokens per query than top-k retrieval, and this approach is sized for a personal-scale vault (hundreds of notes), not millions of documents.
def inject_root_instructions(cache_stable: bool = True) -> str:
# Read fresh on every call (<1ms): human edits in Obsidian are live on the next LLM turn
template = ROOT_INSTRUCTIONS_PATH.read_text(encoding="utf-8")
...
vault_structure_tree = get_directory_tree(path=None, show_files=False, max_depth=2)
# Hour-aligned timestamp so per-second ticks never bust the provider's KV prompt cache
timestamp_str = (
datetime.now().strftime("%Y-%m-%d %H:00")
if cache_stable
else datetime.now().strftime("%Y-%m-%d %H:%M:%S")
)
return (
template
.replace("{user_name}", settings.nexus_user_name)
.replace("{current_time}", timestamp_str)
.replace("{vault_structure}", vault_structure_tree)
.replace("{shared_tools}", get_shared_tools_catalog())
)Recently Shipped
v3.4.0 changelog- •Conversation branching DAG with distill-and-merge back into the parent thread
- •Image & PDF upload with text-layer + table extraction for multimodal chat
- •Read-only Chrome tab-group bridge (MV3 extension → WebSocket → Markdown)
- •Stop & Revert: cancels a running turn and rolls back its checkpoints
- •HITL stale-base detection, 3-way rebase and one-click rollback
- •Mobile PWA control panel over LAN / Tailscale
Tool Modules
12 deterministic Python/Node utilities in src/nexus/shared_tools/, callable by the agent or from the CLI.
Gmail REST API v1, multi-account search, threads, labels and drafts. Never sends.
Google Calendar OAuth2 client for listing, creating and updating events.
Tavily-backed web search with depth/domain filters and Markdown citations.
Reads open Chrome tab groups and rendered DOM via an unpacked MV3 extension.
Extracts clean article Markdown from a URL with trafilatura.
Downloads YouTube transcripts to text for note distillation.
Parses HL7 CDA clinical XML exports into structured Markdown.
ADB screen-scraper for 11 Android chat apps via the UI hierarchy.
Turns a Markdown note into an MP3 with edge-tts.
JSON/JSONL chat transcript → Markdown with turn coalescing.
Pure-Python vault navigation, search, recent-edit queries and PDF ingestion.
Playwright PDF + DOCX resume renderer with page-fill metrics.
End-to-End Execution Pipeline
Click any stage below to inspect the real Python module and function that handles it in Nexus.
Universal NexusAgent & Tiered Model Routing
One universal LangGraph StateGraph (agent ⇄ tools) replaces rigid domain routers. It navigates the Vault directly, reads <Section>/Section Instructions.md on demand, and routes across Gemini, OpenAI, and Anthropic tiers—automatically falling back to a low-latency tier if a primary model hits a 504 deadline timeout.
def call_model(state: NexusAgentState, config: RunnableConfig | None = None) -> dict:
...
system_prompt_content = inject_root_instructions()
ephemeral_sys_msg = SystemMessage(content=system_prompt_content)
messages_for_llm: list[BaseMessage] = [ephemeral_sys_msg] + cleaned_messages
...
active_llm = get_chat_model(tier=requested_tier, temperature=0.0)
try:
response = active_llm.bind_tools(TOOLS).invoke(messages_for_llm)
except Exception as e:
...
if is_deadline_or_timeout and effective_tier != "low":
fallback_llm = get_chat_model(tier=fallback_tier_arg, temperature=0.0)
response = fallback_llm.bind_tools(TOOLS).invoke(messages_for_llm)Professional Experience & Track Record
300+ hours of LLM and multimodal evaluation paired with clinical billing automation and 8 years of business operations and HIPAA-regulated medical records work.
AI Training Fellow
300+ Hrs Model Evaluation- Completed 300+ hours of human evaluation and comparative grading across LLM and image generation outputs; tested prompts for factual errors, hallucinations, and instruction-following failures against structured rubrics.
- Performed iterative resampling with corrective feedback across LLMs and latent diffusion models, identifying systematic failure modes and context degradation patterns.
Python Billing Automation (Contract / Independent)
Clinical Workflow Automation- Built Python scripts for an independent LCSW clinical practice to automate insurance billing workflows and replace manual claim entry, eliminating recurring claim formatting errors and saving weekly administrative hours.
Co-Owner & Operations Lead
2x Revenue Over 18 Mos- Doubled private-brand revenue over 18 months by engineering automated inventory forecasting pipelines, competitor pricing analysis, and catalog SEO across Amazon Seller Central, Shopify, and wholesale channels.
- Streamlined international fulfillment and unit margins across US, UK, and EU hubs by developing automated data reconciliation workflows and SQL reporting for customs logistics, cross-border freight, and supplier replenishment.
Medical Records & Patient Services
100% HIPAA Compliance- Co-managed patient charts and medical records for a high-volume ophthalmology practice, maintaining 100% HIPAA compliance across EHR entry, chart retrieval, and inter-office record transfers.
- Created structured tracking sheets for patient flow and staff scheduling to reduce chart retrieval wait times during peak clinic hours.
Education & Academic Honors
Bachelor of Science in Computer Science
Dean's List. Coursework in Algorithms, Operating Systems, Database Systems, Artificial Intelligence, and Software Engineering.
Associate of Science in Computer Science
Foundation in Object-Oriented Programming (Java/C++), Data Structures, and Discrete Mathematics.
Visual Art & Aesthetic Systems
Engineering and visual art share the same root: spatial reasoning, composition, and taste. A curated collection of original physical paintings, digital compositions, and computational studies.

Trip 1
Chromatic study in layered form, light, and spatial depth—also used as the ambient background canvas in the site hero.

Trip 2
Companion piece exploring high-contrast color movement, organic texture, and structural balance.
The Philosophy of Creative & Technical Synthesis
Visual art builds acute spatial intuition and composition discipline that directly informs system architectures, modular component design, and UI telemetry clarity.
Skills & Where They're Used
Every item below points to a concrete module, project, or role where I used it.
Master Resume
Kept in sync with the Nexus repository and my master Markdown resume.
William Volodarsky
Software & AI Engineer (Full-Stack / Applied AI / Enterprise Systems)
Fort Lee, NJ 07024 • bill@williamvolodarsky.com • williamvolodarsky.com • github.com/brolodarsky • linkedin.com/in/williamvolodarsky
01 // Summary
Software & AI Engineer (Full-Stack / Applied AI / Enterprise Systems) building production web applications, agentic workflows, and internal tooling. Computer Science graduate (BS CS, Magna Cum Laude, Dec 2025) combining modern Python, TypeScript, and SQL engineering with 8+ years of real-world operations and clinical data systems. Builder of Nexus (a LangGraph + FastAPI + Next.js 16 platform with pre-call prompt hydration, parallel subagent swarms, conversation branching, and a two-phase commit HITL queue verified by 171 pytest tests) and MEM Billing (clinical practice automation).
02 // Projects & Engineering
Monaco Editor + hitl.db) with cumulative edit stacking, stale-base conflict detection, 3-way line rebase, and revision rollback.StateGraph (NexusAgent) with hour-aligned pre-call instruction/tree hydration (inject_root_instructions), tiered Gemini / OpenAI / Anthropic routing (504 fallback), and parallel subagents (spawn_subagent) on a SQLite blackboard.search_vault, read_file, get_directory_structure, get_recent_modifications, PyMuPDF text + table extraction) that navigate domain instructions and whole notes on demand.browser_bridge.py), Tavily search, trafilatura webpage/YouTube extractors, an HL7 CDA XML parser, an ADB Android chat scraper (11 apps), and a Playwright PDF + python-docx resume compiler.audit_run_costs.py).Built Python scripts for an independent LCSW clinical practice to automate insurance billing workflows and replace manual claim entry, eliminating recurring claim formatting errors and saving weekly administrative hours.
03 // Professional Experience
Completed 300+ hours of human evaluation and comparative grading across LLM and image generation outputs; tested prompts for factual errors, hallucinations, and instruction-following failures against structured rubrics.
Doubled private-brand revenue over 18 months by engineering automated inventory forecasting pipelines, competitor pricing analysis, and catalog SEO across Amazon Seller Central, Shopify, and wholesale channels. Streamlined international fulfillment and unit margins across US, UK, and EU hubs by developing automated data reconciliation workflows and SQL reporting for customs logistics, cross-border freight, and supplier replenishment.
Co-managed patient charts and medical records for a high-volume ophthalmology practice, maintaining 100% HIPAA compliance across EHR entry, chart retrieval, and inter-office record transfers. Created structured tracking sheets for patient flow and staff scheduling to reduce chart retrieval wait times during peak clinic hours.
04 // Education
Let's Build Production AI Together.
I am currently evaluating high-impact Agentic AI Engineering and Systems Orchestration roles (NYC, Hybrid, or Remote). Whether you are building autonomous agents, tackling LLM alignment, or scaling complex operations, let's connect.