BROWSER-BASED AGENTIC AI SYSTEMS · OVERLIFT · SEAN FINDLEY
I build browser-based agentic AI systems that explore deeply—without taking control.
I’m Sean Findley, creator of OverLift and a senior software engineer who turns ambitious AI ideas into products people can understand, inspect, operate, and trust.
These systems can understand questions, search local and approved evidence, compare options, explain recommendations, and prepare useful next steps. Private work can stay in the browser; exact software limits permissions and official changes; people approve consequential actions; receipts show what happened and why.
THE OVERLIFT PROMISE
Useful intelligence. Deliberately small power.
OverLift gives AI room to search, compare, explain, and propose while private data, permissions, tools, and official decisions remain under explicit human and software control. Intelligence can help without receiving the keys to the system.
-
PRIVACY
Keep information within its chosen boundary.
Browser-local and same-origin intelligence can reduce unnecessary data movement and external disclosure.
-
SECURITY
Limit what every agent can reach and do.
Tools, data, persistence, and outside services remain behind explicit capabilities and policy.
-
AUTHORITY
Control what may become official.
Models can propose. Trusted code and approved people decide what changes state, becomes a record, or triggers an action.
SEE OVERLIFT IN ACTION
One browser demo shows the whole promise.
Ask a plain-language logistics question and watch OverLift find evidence, compare recovery routes, expose uncertainty, stop at permission boundaries, request the exact approval, execute once, verify the result, and save a replayable Decision Receipt.
Finds evidence and shows the source
Stops when the evidence is too weak
Cannot use a tool it was not given
Shows the exact action before approval
Prevents duplicate execution
Verifies the result and saves a receipt
FOR MARKETING, BUSINESS, ARCHITECTURE, AND ENGINEERING
I connect the product promise to the system that has to deliver it.
I can explain the value in plain language, shape the architecture, build the working product, measure its quality and failure behavior, and carry it through delivery without losing the original business goal.
- Marketing Create more relevant experiences without casually exporting customer data, inventing claims, or hiding why a recommendation appeared.
- Business Move research and operations faster while keeping approvals, accountability, duplicate protection, and consequential decisions under control.
- Architecture Keep models, tools, data, permissions, and official state behind explicit, testable boundaries that fail closed.
- Engineering Ship local-first and agentic systems with measurable retrieval quality, deterministic gates, receipts, replay, recovery, and clear failure behavior.
SELECTED WORK
Six products that show how I think and build.
These six show what I have actually built across operations, research, Azure learning, private document assurance, real-time agent behavior, and browser systems. Each is a distinct product with a plain-language purpose, an OverLift advantage, and a focused case study containing the engineering proof.
OverLift Logistics
An SLA exception-intelligence product for shipments, service parts, inventory, couriers, and recovery operations. It assembles fragmented operational evidence, explains why an order is at risk, compares viable recovery routes, and stops at the correct human-approval boundary.
A working logistics operations system where agents investigate a disruption, retrieve local evidence, compare recovery routes, and prepare a bounded plan. Exact software keeps inventory rules, permissions, approvals, duplicate protection, execution, and receipts outside the model’s authority.
Competitive advantage
Most logistics assistants summarize the problem; OverLift Logistics connects time-aware evidence to typed tools, deterministic validation, and post-action receipts. That makes every recommended route inspectable, permission-bound, and reproducible instead of merely plausible.
- Faster exception resolution
- Stronger SLA protection
Saros.ai
A reproducible market-intelligence workstation that reconstructs what was knowable at a chosen point in time, connects price behavior to governed evidence, tests a thesis, and preserves the assumptions, calculations, sources, and receipt behind the result.
An offline research and decision workstation that combines market, regulatory, economic, crypto, and news evidence across time. It preserves what was actually knowable, exposes disagreement and exclusions, and can reproduce the research conclusion later.
Competitive advantage
Most market tools mix current knowledge with historical analysis and make hindsight leakage hard to detect. Saros freezes the as-of evidence, keeps semantic organization separate from deterministic math, and makes a thesis replayable so later conclusions can be compared honestly.
- More defensible research
- Faster thesis testing
AzureGlossary
A connected Azure learning and reference platform with more than 2,000 terms, 5,000-plus CLI references, reviewed certification tracks, official-source links, semantic search, and guided operational tools. It helps learners move from the question they actually have to the right concept, command, source, or study step.
A practical Azure and agentic-AI learning system with semantic search that handles natural questions, synonyms, and misspellings. It returns grounded terms, relationships, study paths, and measurable retrieval quality instead of relying on a brittle keyword box.
Competitive advantage
Instead of forcing users to choose between fragmented documentation, exam cramming, and a generic chatbot, AzureGlossary connects vocabulary, commands, objectives, and official evidence in one deterministic knowledge graph. Semantic search improves discovery while Microsoft documentation remains the authority.
- Clearer Azure learning paths
- Faster concept lookup
Vellucent Studio
A private-by-default PDF assurance workspace for sensitive, regulated, and government documents. It combines page-aware evidence, exact and semantic retrieval, proposal-only grammar agents, Meaning Lock, human review, verified derivatives, and receipts so better writing never outranks the record.
A private PDF document-assurance system that reconstructs page-aware language, proposes bounded grammar corrections, and preserves the original record. Meaning Lock, deterministic admission, human review, and receipts prevent a fluent edit from silently changing facts, policy, or intent.
Competitive advantage
Generic PDF copilots flatten structure and treat fluent output as success. Vellucent retains page/object anchors, separates exact evidence from learned interpretation, measures required corrections and KEEP cases, protects consequential meaning, refuses unsafe changes, and preserves a reversible proof of every decision.
- Lower document exposure and cloud dependency
- Traceable, reversible corrections for government and regulated review
Neon Drift
A neon arcade game and the completed two-agent reference implementation of the OverLift NPC Agent Fabric. Zarvox watches and directs from above the arena; VANTA-9 is the physical red hunter inside it. They communicate, disagree, learn player patterns, and change behavior while the game world retains final authority.
A real-time two-agent game-intelligence demonstration where VANTA-9 and Zarvox plan, adapt, speak, and react to actual play. It proves bounded agent behavior, deterministic game authority, browser performance, and explainable NPC decisions inside one visible system.
Competitive advantage
Most 'AI NPC' demos showcase generated dialogue. Neon Drift proves governed physical behavior: two independent agents can accept, modify, delay, reject, or abstain; semantic memory detects conceptual repetition; Q-Lens evaluates consequences; and a deterministic compiler prevents stale, unfair, or impossible actions.
- Less repetitive NPC behavior
- Buyer-grade explainability and replay
Vectrexia
A polished Vectrex emulator and digital museum that combines accurate vector-machine behavior with a curated cartridge library, speech-routed commands, local intent handling, diagnostics, and a consumer-friendly browsing experience.
A browser-native vector arcade laboratory that combines exact machine behavior with a curated, approachable product experience. Speech routing, private diagnostics, deterministic state, and repeatable compatibility evidence show how OverLift patterns apply to demanding real-time software.
Competitive advantage
Generic emulator front ends often begin with a file picker and leave compatibility work hidden. Vectrexia treats the original library as a designed product, makes speech and local intent bounded by deterministic machine state, and preserves private diagnostics that make failures reproducible.
- Faster emulator navigation
- Higher emulator trust
HOW OVERLIFT BECOMES A PRODUCT
A practical method for useful, controlled AI.
Start with the real problem, let intelligence explore within explicit boundaries, then let trusted software and people approve, execute, verify, and prove the result.
Ground the real problem
Clarify the goal, evidence, time, uncertainty, and who is allowed to decide. Missing or conflicting facts stay visible.
Let intelligence explore safely
Search, compare, and propose without allowing a model or document to rewrite facts, expand permissions, or bypass policy.
Commit and prove the result
Trusted code performs only the approved change, verifies the outcome, and preserves the evidence, approval, state transition, and receipt.
BRING ME THE DIFFICULT PROBLEM
Turn an ambitious AI idea into something useful, controlled, and ready to ship.
I work across product judgment, architecture, implementation, validation, and delivery—especially when the workflow is ambiguous, the evidence matters, or uncontrolled AI behavior is not acceptable.