Divya Does AI · A recurring experiment

Quarterly AI Review.

What I asked AI to do. What actually happened. What broke. And was it worth managing?

A candid, evolving assessment of using AI as an execution partner across building, writing, research and everyday operations. Not a productivity victory lap.

By Divya Kandikatti · Covers Q3 and Q4 2026 · Published · Updated

THE BOARD READOUTDecision brief · qualitative

The central finding

AI accelerates the start.
Delivery is the constraint.

Our experiments show creative and research leverage, but a recurring gap between useful output and independently verified completion.

01Where value appearsIdeation · research · writing
02Where confidence fallsQA · completion · context errors
03What needs evidenceTime saved · ROI · first-pass success
Executive snapshot / Q3Curated examples only · No measured ROI
Experiments featured04Not total tasks
Observed examples02Not externally audited
Verification pending02Evidence gap
Net time saved—Baseline not recorded

Engineering reality / Sep–Nov 2026

The numbers behind the experiments.

Actual GitHub and Vercel activity, not AI self-ratings. These counts represent engineering activity, not independently verified production delivery.

Vercel deployments250Returned in the Sep–Nov query
Ready211Built successfully; QA not implied
Failed20Deployment state: ERROR
Canceled19Deployment state: CANCELED
September 202648Pull requests created
October 202629Pull requests created
77 PRs across two months.

Real delivery activity, but the next question is how many of those PRs produced verified outcomes versus rework.

Source: connected GitHub PR search by creation date (September 1–October 31, 2026) and Vercel deployment list for the main website project queried from September 1 through November 2, 2026. The Vercel query returned 250 records, with the oldest returned deployment on September 14; this is a bounded activity window and must not be interpreted as full Q3 or Q4 totals. Ready ≠ QA passed. Other repositories and projects are not included. Source records were reviewed through authorized private connections; the repository and deployment logs are not publicly accessible. These figures should be treated as a dated research snapshot, not independently public-auditable data.

Transparency before impressive numbers. The ratings below are a qualitative retrospective, not measured KPIs. The experiment cards are curated examples, not a complete activity count. Quantitative outcomes will appear only after source validation.

01 / Collaboration

The performance review

Provisional · retrospective

The first scorecard reflects our conversation-based self-assessment across the wider August–November collaboration. It is not a Q3-only measurement, and should not be compared numerically across quarters yet.

Collaboration capability profileSeven qualitative dimensions · 0–10
The execution disconnectJudgment-based assessment / 10
Key observationIdeation: 9/10. End-to-end completion: 4/10.

Our principal gap isn't generating solutions. It's finishing independently with trustworthy evidence.

The pattern: ideation and content are stronger than independent QA and reliable end-to-end completion. The scores are editorial judgments; detailed source-backed KPIs are a future iteration.

02 / Portfolio

What we actually tried

Jul 1 – Sep 30
Evidence confidenceSelected examples, not audited delivery rate
2 / 4

Examples with an observed outcome

Observed examples2
Needs verification2

Visibility is not verification. We keep unverified and incomplete work visible rather than treating every activity as success.

Where AI showed upQ3 · Select a domain to filter the experiments below

A map of selected workstreams, not a claim about time spent or total volume. Tap a category to explore the underlying examples.

BuildObserved

Personal website and interactive professional story

A rough creative direction can become a tangible web experience, but verification still matters.

Explore related work
LeadNeeds verification

Project Odin and agentic delivery exploration

Professional examples require careful confidentiality and source confirmation.

CreateObserved

AI-assisted writing and podcast experiments

AI helps develop structure; authentic voice needs deliberate human editing.

Explore related work
OperateNeeds verification

Long-running research and decision support

Continuity is useful, but remembered information must not substitute for verified facts.

Examples are selected from our collaboration and may span longer workstreams. Status is conservative: “Observed” does not mean end-to-end externally audited. Private correspondence and client details are intentionally omitted.

03 / Measurement

What would prove value?

Baseline pending

Instead of inventing time-savings, success rates or ROI, here's the measurement contract for the next review.

Not yet measured

Verified outcomes

Deliverables with evidence and an accepted result

Unit: Count
Not yet measured

First-pass success

Completed tasks accepted without corrective rework / audited completed tasks

Unit: %
Not yet measured

Human interventions

Times the human had to correct, unblock or re-explain

Unit: Count
Not yet measured

Net time saved

Manual baseline effort minus AI supervision, retries and rework

Unit: Hours
Not yet measured

QA reliability

Valid checks passed on the reviewed commit / checks executed

Unit: %
Not yet measured

Deployment efficiency

Verified delivered changes / deployments

Unit: Ratio
Not yet measured

Context accuracy

Correct audited personal facts / all audited personal facts

Unit: %
Not yet measured

Cost per outcome

AI and supporting tool spend / verified outcomes

Unit: USD

Net value, not just output

Useful work delivered − human supervision and rework − tool costs. Time and money should be reported in their own units until a defensible valuation is established.

04 / Retrospective

Start. Stop. Continue.

Start

Measure outcomes, not activity

Collect evidence, interventions and rework before claiming time savings.

Start

Define done before implementation

Differentiate built, deployed, tested, approved and published.

Stop

Automating broken automation

Investigate root causes before adding another workflow.

Stop

Treating remembered facts as verified

Cross-check important personal, technical and public statements.

Continue

Turning rough ideas into tangible work

Prototype, build and iterate using the existing stack.

Continue

Human approval for consequential changes

The human retains publishing and production decisions.

The question for next quarter

Am I spending more time doing what I care about—or managing AI?

This dashboard is the evidence companion to a future first-person Substack story. Publication is pending editorial review.