FLAGSHIP · PERSONAL LAB
The Brain
A private knowledge layer that keeps provenance, structure and controlled updates visible throughout retrieval.
KnowledgeRetrievalProvenance
Explore The Brain →AI SYSTEMS · AUTOMATION · ANALYTICS
I turn messy workflows into working systems.
I design and build practical AI, automation and data products — from governed multi-agent systems to 3D spatial tools and operational decision workflows.
5+ yearsdata → decisions
8 systemson this site
TorontoCanada
What I build
Systems that make complex work easier to understand, operate and trust.
01Applied AI
Agents, retrieval, extraction, reasoning and human-in-the-loop workflows.
02Automation
Repeatable workflows that reduce manual coordination and keep state visible.
03Data products
Decision tools, interactive dashboards and business-facing analytics.
Selected systems
View all work →See the idea in seconds. Explore the system when you want depth.
FEATURED · PROFESSIONAL WORK
Spatial Intelligence
Two paths — public imagery or architectural plans — converge into validated 3D building models for an operational platform.
Computer Vision3DDeterministic Geometry
Explore Spatial Intelligence →Operating principle
“Use AI where reasoning helps. Keep authority, evidence and deterministic mechanics explicit.”
That principle connects the work across agents, knowledge systems, analytics and spatial tooling.
Selected work
AI, automation and decision systems.
A curated set of systems spanning personal R&D and professional work. Company work is intentionally generalized and demonstrated only with synthetic data.
01
Flagship systems
02
Professional case studies
03
Selected builds
06
Receipt Intelligence
Structured extraction and ledger workflow from unstructured receipts.
OCRAIWorkflow
07
Restaurant Operations
Operational planning, finance and actual-vs-plan workflows.
PlanningFinanceAnalytics
08
Insurance Workflows
AI-assisted quotation, remittance and structured operations tooling.
AI AssistWorkflowCRM
← Work
Hermes
Governed Multi-Agent AI Operating System
A control architecture for AI workers where authority, routing, verification, evidence and repair are explicit system concerns.
Multi-AgentGovernanceSQLiteVerification
The idea
Agents are workers. Governance is the system.
Hermes separates human authority, deterministic control, model routing and verification so useful AI output does not automatically become trusted action.
The current implementation includes a canonical authority core and protected-effect primitives; some routing, reporting, repair and continuity paths remain staged, tested or blocked rather than universally live.
01
Authority as data
Ownership and permission are explicit system state, not persona text inside prompts.
02
Role ≠ model route
Who is responsible, how risky the task is and which model class executes it are separate decisions.
03
Independent verification
Authors do not self-approve protected results.
04
Bounded repair
Failures create new attempts while preserving the evidence that came before.
Task state READYrepair round 0/3
01 · Authority
02 · Control plane
03 · Delivery
04 · Repair
flow · authorityadvisoryrepair
model routeD0rulesM1fastM2standardH3strongest
Authority became data, not persona prose.
Moved critical ownership and protected effects into explicit state rather than relying on prompts to behave correctly.
Verification was bound to evidence.
Release logic checks current artifacts and review ordering instead of accepting a success narrative.
Repair preserves failure history.
Each fix is a new bounded attempt, keeping negative evidence visible rather than overwriting it.
Deterministic mechanics stay out of LLM orchestration.
Counts, scheduling, projections and control-plane rules remain software concerns where possible.
← Work
The Brain
Private AI Knowledge & Intelligence Layer
A local-first knowledge system designed around provenance-bearing Markdown, structured retrieval and controlled change.
Markdown/GitSQLite FTS5RetrievalProvenance
The idea
Knowledge should stay traceable after AI touches it.
The Brain keeps source context, structure and provenance visible from ingestion through retrieval. It uses taxonomy and linked-note structure before falling back to full-text search.
Core local retrieval is implemented and tested. Governed write-back and broader local intelligence remain in development.
42offline tests passed
Local-firsthuman-editable knowledge
Lexical-firstSQLite FTS5, no embeddings yet
01
Ingestscreen sources and preserve identity
02
Structureparse metadata, links and note taxonomy
03
Retrievestructure-first, then lexical fallback
04
Citereturn context with provenance
05
Propose updatepreview / confirm before trust changes
Structure first, search second.
Taxonomy, maps-of-content and linked notes are used before full-text fallback so retrieval respects how knowledge is organized.
Provenance is part of the object.
Source context survives ingestion instead of becoming anonymous text.
Knowledge proposals are not trusted knowledge.
Write-back is designed as a candidate → validate → commit lifecycle rather than silent self-editing.
Complexity is earned.
Local lexical retrieval was implemented before adding embedding complexity that was not yet needed.
← Work
Spatial Intelligence
2D inputs → validated 3D building models
Two complementary workflows create 3D building models when the source material is either public imagery or an architectural floor plan.
Computer VisionThree.jsDeterministic Geometry3D
Public demo noteAll visuals, addresses, floor plans and property data on this page are synthetic. Company details are intentionally generalized.
The system
Two ways in. One usable 3D building model out.
When a floor plan is unavailable, the workflow derives geometry from public imagery. When an architectural plan exists, the alternate path extracts walls and floor structure directly from the drawing.
Both paths were built to support 3D building use in an operational platform. The imagery-based workflow has been used across roughly 30 buildings before operational automation was handed over to IT with guardrails.
PATH A · NO FLOOR PLAN
Public imagery
Locate → trace → validate → estimate floors → build 3D.
+
PATH B · FLOOR PLAN AVAILABLE
Architectural PDF
Extract vector structure → identify walls → stack floors → build 3D.
→
3D building model
Ready for integration into the operational platform.
Breakthroughs01 02 03 04
AI stopped owning geometry.
The workflow evolved from model-generated shapes toward measured, deterministic extraction and explicit validation.
Testing moved onto the shipped path.
A reimplemented test harness hid real bugs, so validation was redesigned to exercise the actual code path and fail loudly.
Random segmentation became deterministic tracing.
A pure-JS footprint tracer replaced a randomness-heavy dependency to make repeated inputs stable.
Weak evidence is refused, not polished.
Lock gates were redesigned around the outline actually used, with uncertain geometry surfaced rather than silently accepted.
← Work
Account Health
Proactive analytics & decision support
A business intelligence experience that brings operational signals into a clearer view so teams can identify attention areas, investigate patterns and take proactive action.
Power BIAutomationDataDecision Support
Public demo noteThis case study intentionally stays broad. All accounts, values and operational signals shown are synthetic.
The shift
From static reporting to proactive account review.
The project began with user discovery and a redesign of how account information is reviewed. The final experience consolidates relevant signals into layered views, supports drill-down investigation and automates recurring history so teams can move from “what happened?” toward “what needs attention next?”.
It is used as an operational decision-support tool by the team and leadership. Public details are deliberately generalized to protect internal business logic.
01
Listen before rebuilding
User interviews changed the product direction and clarified the real job the dashboard needed to do.
02
Layer complexity
Information is separated into navigable levels rather than forced onto one crowded surface.
03
Make trends operational
Automated historical snapshots make patterns easier to review over time.
04
Design for action
The experience emphasizes attention signals, investigation and proactive follow-up rather than passive reporting.
Breakthroughs01 02 03 04
The dashboard became a decision system.
The design shifted from presenting metrics to supporting a repeatable scan → investigate → act workflow.
User behavior reshaped the information architecture.
Discovery showed how people actually used the tool, leading to layered views and clearer client-facing presentation.
Historical context became automated.
Recurring snapshots moved trend review from a manual idea into a maintainable workflow.
Clarity builds trust.
Definitions and explanation were placed closer to the point of use so people could understand what they were seeing while they worked.
← Work
Agent Office
Role-based multi-agent operations workspace
A multi-agent workspace that assigns specialist roles, coordinates work, cross-checks outputs and produces a consolidated summary for human review.
AgentsOrchestrationVerificationHuman Review
Current boundary
Useful autonomy, visible limits.
The current workflow can run through specialist work and generate a consolidated summary for human review. It is intentionally not presented as fully autonomous production execution.
The strongest engineering theme is reliability: failures are surfaced, agent cooperation is throttled around platform limits and human review remains the final control point.
Breakthroughs01 02 03 04
Silent simulation was removed.
One early failure could latch the workspace into canned responses that looked real. The fix made live/degraded state explicit and recoverable.
The runtime mattered more than the retry logic.
Moving to the correct execution environment unlocked real model calls without credential hacks or a rewrite.
Live model behavior exposed parser assumptions.
Structured handoffs were redesigned so routing did not depend on a directive appearing at a fixed position in model output.
Reliability beat maximum concurrency.
Parallel orchestration was reduced in favor of sequential, throttled phases with retries and verification.
About / experience
Builder, analyst and operator.
My work sits between business problems and technical systems: understand the workflow, find the friction, build the tool, test it with real users, then make it repeatable.
5+ years turning client and operational data into decisions — from KPI dashboards and client-facing reporting to production AI tools that colleagues use every day.
Experience
Lead Analyst, Client Success
Canadian utility submetering company
Leads analytics, decision-support tooling, workflow automation, AI-enabled prototypes, team enablement and strategic problem solving. Work spans Power BI, Power Automate, APIs, data analysis, process redesign and applied AI.
- Builds the SQL- and Power BI-based KPI dashboards leadership uses, retiring manual Excel reporting.
- Designed an audit process end-to-end, from diagnosis to rollout, with automated workflows and an interactive dashboard.
- Shipped a suite of sanctioned internal AI tools used daily by colleagues — a team assistant, a PDF data-extraction tool, a 3D building visualization tool and workflow automations.
- Built an LLM-assisted ticket triage agent that pulls context and returns structured summaries for human action.
Data & insightAutomationAI initiativesTraining & guidance
Client Success Analyst
Canadian utility submetering company
Produced client-facing performance insights for account reviews with property management clients, partnered across teams to validate data and map current-to-future workflows, and helped streamline reporting and operational handoffs.
Data Validator
Canadian utility submetering company
Validated large operational datasets, resolved discrepancies, documented root causes and recommended controls to prevent recurrence.
Data Analyst
Think Insurance (insurance broker)
Built management reporting, digitized office workflows — cutting time spent on manual data entry by 30% — and led a Zoho CRM rollout, including workflow configuration, documentation and team training.