LastLook AI

Inspiration

Construction has no undo button. Some of the most consequential quality checks happen in a short window immediately before work becomes concealed: rebar before concrete, MEP services before drywall, fire stopping before finishes, and waterproofing before the next layer. Today, the worker sees the real condition while the relevant drawings, specifications, and checklists are often somewhere else. Evidence is usually a handful of photos, a phone call, and memory.

Kaya AI's Physical AI brief asks for real-time, hands-free intelligence at the point of work. LastLook AI focuses that intelligence on the most irreversible moment: the last inspection before work disappears.

What it does

A worker wearing smart glasses looks at the installation and asks, "LastLook, is this bay ready to pour?"

LastLook:

1. Grounds the first-person view in the correct location, drawing, specification, and inspection checklist.
2. Uses a vision-language model to compare the visible condition with the retrieved requirement.
3. Requests a closer or second view when evidence is incomplete, and abstains when confidence is low.
4. Creates a marked issue if something looks wrong and attaches the exact drawing or specification reference.
5. Stores before-and-after evidence and routes sign-off to the licensed QA engineer.

The human remains the authority for structural and code compliance. LastLook does not auto-certify work; it makes retrieval, anomaly cues, and evidence completeness hands-free and immediate.

Why this is different

Most construction computer vision products focus on progress tracking, broad safety alerts, or retrospective documentation. LastLook owns a narrow, high-value checkpoint where a missed defect is about to become hidden and expensive.

That wedge creates three advantages:

• Immediate ROI: one avoided concealed-work defect can justify deployment.
• Natural hardware fit: smart glasses see exactly what the worker sees without occupying their hands.
• A compounding moat: every human-approved checkpoint becomes part of a spatial evidence graph linking place, requirement, visual condition, decision, and outcome.

How we will build it

The Stage 2 prototype will combine:

• Smart-glasses video, voice, and timestamps
• Tag-based spatial grounding for the first demo, with BIM-informed visual localization as the next step
• A vision-language model for visible-condition reasoning
• Construction-specific document RAG over drawings, specifications, and QA checklists
• A confidence gate that can ask for another view or abstain
• An action agent that drafts an issue, RFI, or evidence pack for human approval

Our demo will cover one complete pre-pour workflow: first-person capture, grounded check, detected mismatch, correction, human sign-off, and a generated PDF evidence pack.

Market-ready path

We start with concrete contractors and general-contractor quality teams because pre-pour inspections are frequent, structured, and tied to irreversible work. The commercial model is a per-site subscription with active-glasses seats and an enterprise evidence archive.

After proving pre-pour QA, the same concealment-checkpoint engine expands to waterproofing, MEP before drywall, fire stopping, facade anchors, and handover digital twins.

Research literature reports average rework costs around 5% of construction cost. On an illustrative Rs 100 crore project, preventing only 5% of that rework protects Rs 25 lakh. This is a framing for the opportunity, not a performance claim.

Challenges

Construction drawings and site geometry can be messy, and confident-looking model errors are unacceptable in a safety-critical environment. We address this by starting with explicit location tags, narrowing the checklist, requiring confirming views, exposing evidence and confidence, and preserving licensed-human authority.

What we learned

The best Physical AI product is not the one that talks the most. It is the one that understands when evidence is insufficient, retrieves the right requirement, and helps a human act before the physical world makes the mistake permanent.

Team

• Sudhama JVS — IITM
• Pranathi Jayanthi — HP Inc. (working professional)

Built With

• vision-language-models