Product requirements & architecture

What we are building, and the rules it lives by.

Make every camera in every factory answerable in plain language, so that any question about safety or operations can be asked, answered, and acted on in seconds, with the video as proof.

North Star metric

Grounded Answered Questions per Active Site per Week

GAQ / site / week

It only goes up when the product is trusted enough to be asked, answers are good enough to act on, and the habit has formed across personas. Supporting inputs: retrieval precision, answer-faithfulness, time-to-answer.

Strategic pillars

Four non-negotiables

Ask, don't configure

The unit of interaction is a question, not a detector setup.

Grounded or silent

Every answer cites the clip and timestamp. If we cannot ground it, we say so. No hallucinated incidents on a safety product, ever.

Edge-first, privacy-by-design

Raw video stays on site. Only anonymized events and embeddings move. A feature and a moat, not a constraint.

Two budgets, one product

Safety to land, operations to expand. Both run on the same event layer.

What we are explicitly not

Not a camera or hardware vendor. We ride existing RTSP/ONVIF estates.

Not a real-time, safety-rated machine-control system in v1.

Not a worker-performance surveillance tool. We measure processes and hazards, not individuals.

Architecture · VLM + RAG-over-events

Extract events once. Query them many times.

On-site (edge · raw video never leaves)Query-time (only events + embeddings)- extract events once at the edge, query them many times.

Ingestion (edge)

Pull camera streams via RTSP/ONVIF into an on-site edge node. No footage leaves the site.

Perception (edge)

A VLM plus specialized detectors convert video into typed, timestamped events. Faces and bodies are blurred here; identity is never stored.

Event store + index

A vector index for semantic retrieval and a structured event DB for precise filtering. Clips stay encrypted locally; only pointers are indexed.

Retrieval (hybrid RAG)

A question becomes a semantic query plus structured filters (time, zone, event type, shift). Retrieval returns the relevant event set.

Reasoning + generation

An LLM composes counts, trends, and narrative, each grounded to events with timestamp and clip citations. Empty or low-confidence returns 'no supporting footage found.'

Evaluation (continuous)

An offline + online evals harness scores retrieval precision/recall and answer faithfulness, with human-in-the-loop labeling. The trust engine, not an afterthought.

On-site · edge · no egress

Query-time · events + embeddings only

Evaluation (step 06) wraps the entire loop - scores every answer for faithfulness, continuously

Build vs. buy

Perception / embeddingsBuyCommodity-trending VLM (Twelve Labs / Bedrock); not our moat

Manufacturing event ontologyBuildCore domain moat

Edge runtime + anonymizationBuildTrust moat; regulatory necessity

Vector + event store / retrievalBuy infra, build glueStandard infra, proprietary orchestration

NL query + reasoning + citationsBuildThis is the product surface

Evals harnessBuildThe thing that makes answers trustworthy

Integrations (MES/ERP/EHS)BuildStickiness + expansion

Problem statement

Video is the richest record on the floor, and it is write-only.

Today

Scrub hours of footage by hand after an incident
Or buy a rigid single-purpose detector per question
EHS can’t prove programs work; ops can’t trace losses

With FactoryLens

Ask in plain language - no detector to configure
Get a cited, timestamped answer in seconds
Every claim clickable back to the exact clip

Manufacturers operate large camera estates but cannot interrogate them. After an incident or a slowdown, finding the relevant footage means scrubbing hours of video; preventing recurrence means buying single-purpose detectors. Video, the richest record of what happens on the floor, is effectively write-only. EHS teams cannot prove their programs work; operations teams cannot trace losses to root cause; both pay in incidents, downtime, and liability.

Goals

Measurable outcomes

Collapse incident investigation time by 80%

from hours of review to a sub-minute grounded answer

Reduce recordable safety events 25%+ in 6 months

via proactive querying and trend surfacing

Surface 3+ operational loss sources per site in 90 days

micro-stoppages, idle-after-break, bottlenecks, dollarized

Reach 30+ grounded answered questions / site / week

the North Star, within one quarter of go-live

Pass EU works-council / AI-Act review with no raw-video egress

100% of EU deployments

Requirements

MoSCoW with acceptance criteria

Must-have · P0

P0.1Camera ingestion (RTSP/ONVIF)

Events extract continuously without footage leaving the site.

P0.2Manufacturing event extraction

PPE, LOTO zone, restricted-zone breach, forklift proximity, spill, idle/micro-stoppage, cycle complete, as typed timestamped events.

P0.3Natural-language query, grounded answers

Counts/trends/narrative grounded to clips with timestamps; 'no supporting footage found' when empty, never fabricated.

P0.4Clip + timestamp citation + playback

Clicking a citation plays the exact clip from the cited timestamp.

P0.5Worker anonymization by default

Faces/bodies blurred; no biometric identity stored; no disable toggle in regulated regions.

P0.6Role-based access + audit log

Access scoped by role; every query and view logged immutably.

P0.7Evals harness (faithfulness + retrieval)

Answers below the faithfulness threshold are suppressed or flagged.

Should-have · P1

P1.1Weekly safety + ops digests

Quantified trends with clip evidence.

P1.2Proactive pattern alerts

High-risk recurring behaviors, tuned for low false positives.

P1.3Dollarized operational impact

Loss ranking by cost (micro-stoppage, idle).

P1.4EHS/MES read-integration

Pull shift/line/work-order context to enrich answers.

P1.5Multi-site rollup

Ask across plants; compare sites.

P1.6Exportable DPIA / EU-AI-Act pack

Clears procurement and regulator review.

Could-have · P2

P2.1Custom event by natural language

Teach a new event type from a description + examples.

P2.2Near-real-time notifications

Seconds-latency for the highest-severity hazards.

P2.3Agentic workflows

Auto-open a corrective action with owner + deadline.

P2.4Closed-loop ops triggers

PLC-aware, non-safety-rated.

P2.5Air-gapped deployment

For the most sensitive accounts.

Non-goals

Real-time safety-rated machine control / e-stop. different reliability class and liability regime

Individual worker productivity scoring. legally radioactive under the EU AI Act; off-mission

Selling cameras / hardware. drags margins, lengthens sales cycles

Generic cross-industry video search. Twelve Labs serves the horizontal market; depth is our moat

Defect / quality inspection at line speed. crowded, different latency profile; later adjacency

Regulatory & compliance

Hard constraints that gate enterprise + EU deals

EU AI Act

Workplace emotion recognition is banned (Feb 2025); worker performance-monitoring AI is high-risk with full obligations from Aug 2026. FactoryLens is designed and documented as safety/process intelligence, with human oversight and transparency, and avoids any individual performance-monitoring feature.

GDPR + national law

Video of workers is personal data: lawful basis, minimization, purpose limitation. Germany (BetrVG), Italy (Art. 4) and others require works-council consultation before deployment.

Product implications

Default worker anonymization, no audio, role-based access, immutable audit logs, configurable retention and residency, an exportable DPIA/AI-Act pack, and a purpose lock that scopes queries to safety/ops.