In development
In development · pre-pilot

Watch the work. Write the script that does it next time. You review the diff.

In development. We're building this on ourselves first; nobody is using it yet — not even us.

work-lens
Activity-triggered capture
Cost-aware understanding funnel
Deterministic automations, human-approved
Per-tenant privacy seam (in design)
The problem

Back-office teams lose hours to invisible work: copying a customer detail from WhatsApp into an email, then a document, then the CRM. Re-keying the same renewal across a vendor portal and a spreadsheet. The work is repetitive, cross-app and high-friction — and because it lives in the gaps between tools, nobody can see where the time goes. The tools we've looked at either map the work (enterprise-priced) or monitor the worker (SMB-priced surveillance). We haven't found one that does both halves.

How it works

How it will work

01

Capture

A visible desktop agent records activity-triggered screen video (efficient H.265) plus lightweight telemetry — active window, app, idle state, keystroke counts, never contents. A local ring buffer uploads then deletes, so a dropped connection never loses footage.

02

Reduce

Perceptual-hash deduplication drops idle and duplicate frames; only scene-change keyframes survive. Most of a recorded day is repetition, so this is where the volume collapses.

03

Understand

The day is segmented from cheap telemetry first; OCR reads a representative frame per segment, and a vision model labels each segment (not every frame) with what was happening — which is what keeps the cost low.

04

Decide

Repeated sequences are clustered across days and ranked by frequency × time-spent, so the most expensive busywork rises to the top.

05

Automate

For the top candidates, AI drafts a deterministic automation — preferring API/AJAX over browser codegen over computer-use — which a human reviews on a dry-run/diff before anything goes live.

What's inside

Inside WorkLens.

Activity-triggered capture

The agent starts and pauses on real activity with idle auto-pause, instead of grinding through a fixed 9-to-5 window — less noise, less storage. (Built, spike-validated.)

Cost-aware understanding funnel

Dedup → segment → OCR a representative frame → label the segment. Feeding the model the reduced set, not raw video, is the difference between a low-tens-of-dollars and a four-figure monthly bill. (Code-complete; live-labelling gate pending.)

Deterministic automations, human-approved

Candidates become reproducible, auditable jobs a person signs off on a diff — not an autonomous agent clicking through your billing or a government portal. (Designed; pending real multi-day data.)

Per-tenant privacy seam (in design)

Consent capture, configurable retention, PII redaction, audit log and RBAC as the planned boundary — architected as a first-class seam, not yet implemented end-to-end.

Honest dev milestones

What's actually proven so far.

We're deliberate about what's proven versus planned. These are development milestones on our own data — not customer benchmarks.

  • Spike: screen capture ≈ 33 MB/hour (H.265 at 5 fps) on low-motion content — far cheaper to store than expected. (Single-machine spike, not a production figure.)
  • Spike: OCR remained legible from compressed frames — a dense payment table was read correctly off one compressed frame, confirming the reduce-then-read approach.
  • The full pipeline is code-complete end-to-end (capture → reduce → OCR → segment → label → mine → report). What remains needs a real rollout.
Who it's for

Who actually uses this.

SMBs with repetitive back-office work

Teams doing cross-app data entry — renewals, document prep, vendor coordination — who can't justify enterprise task-mining and don't want a toxic monitoring dashboard.

“Move the same info between five apps” teams

If your team's day is shuttling information between WhatsApp, email, documents, portals and a CRM, WorkLens is being built for exactly that.

Under the hood

The stack

H.265 screen captureLocal-first ring bufferPerceptual-hash dedupOCRVision LLM segment labellingSequence miningDeterministic automationPer-tenant privacy seam
Integrations
  • The desktop agent. A visible, consent-first capture agent on the staff machine — never covert. Employees can see their own data.
  • Automation targets. Drafted automations target the CRM and web apps the work already happens in (API/AJAX first, browser codegen as fallback).
FAQ

Straight answers.

Is WorkLens available yet?+

No. It's in development, pre-pilot — we're building it on our own operation first. Early access list is open; we're not selling licences yet, we'll talk when WorkLens is closer to ready.

Is this employee-monitoring / bossware?+

No, and it's designed to be the opposite. The agent is visible (never covert), it captures keystroke counts but never contents, and its output is automations that remove work — not surveillance scores or disciplinary reports.

How does consent and privacy work?+

The agent is visible to staff, and the product is being built with a per-tenant privacy seam: consent capture, configurable retention, PII redaction, an audit log and role-based access — DPDP-aware from day one, not yet implemented end-to-end.

Will it automate things without us approving them?+

No. WorkLens drafts deterministic automations and a human reviews each one on a dry-run/diff before it ever runs. Nothing touches your live systems unattended.

In development

Where this honestly stands

In development / pre-pilot. Spike + code-complete-pipeline stage with zero production rollout. Froshtek is its own customer zero — there is no reference deployment. Spike numbers shown are from limited single-machine sampling and are labelled as such. This is not “proven in production,” and is not yet generally available.

Want WorkLens early?

Early access list. We're not selling licences yet — we'll talk when it's closer to ready.

30 minutes · no obligation · we reply within 1 business day