Adit Manurkar

AI enablement · business transformation

I turn AI from a side project into how the work gets done.

Ten years delivering AI, analytics, and automation inside planning, engineering, and operations teams at Amazon, Cummins, and Marelli. I take a use case from business problem to deployed tool: scoping it, building the first working version myself, validating it against the current process, and teaching the team to own it.

200+hours a year removed by one AI reporting pipeline
4teammates now building with AI on their own
~$1Mnot spent on an AI program the data could not support
4global teams moved off a decades-old manual workflow

Amazon  ·  Cummins  ·  Marelli  ·  Tesla

Selected work

Six programs where the job was not to use a new technology, but to get it adopted, trusted, and kept after I stepped back.

01

Amazon · Last Mile carrier mix planning · 2025 to 2026

AI-written variance narratives for a weekly leadership review

2-3 hours of manual work every Monday, down to about five minutes

  • Narrative generation
  • Variance analysis
  • Management reporting
  • Human-in-the-loop
  • Baseline benchmarking

The situation

Every Monday, leadership reviews a carrier mix and rate bridge for the Canada delivery network: a variance bridge explaining what moved week over week, and why, through mix and rate effects. It shapes cost and capacity decisions, so it has to be accurate, on time, and consistent week to week.

Producing it took 2 to 3 hours across six manual steps: run the queries, export, reformat into a template, interpret the variances, write the narrative, paste it into the review document. The work produced no new insight, it just moved data. During a major sales event it became a daily race against an 8:00 AM deadline. There was no tooling team, no analyst allocation, and no budget.

What I did

  1. Documented before automating. I captured the data sources, query logic, output templates, and a set of historical bridges, so an agent could execute the workflow and match the tone and structure leadership already trusted.
  2. Used history as the quality bar. Variance interpretation is judgment, not formatting. I reviewed early outputs against past bridges and kept a human check in place until the interpretation held up on its own.
  3. Shipped it unattended, in stages. The risk I cared about was the pipeline quietly putting a wrong number into a leadership document, so trust was earned rather than assumed: validation encoded into the workflow, outputs anchored against historical bridges, and a review period before any unattended run. The daily cadence during the sales event was the hardest live test the quality bar could face.

Outcome

Weekly preparation time
2-3 hrs to ~5 min
Reduction in manual effort
95%+
Daily bridges during a major sales event
15-20 min
Tooling budget and added headcount
$0
Annualized hours returned to the team
200+

Delivery network partners who used to request performance reads one at a time now get the same standardized package every week, which also removed the week-to-week variation in how carrier performance was reported.

Most people start automation by writing code. I started by writing down exactly how a good bridge is made, with examples of good ones. The bottleneck to automating knowledge work is documentation, not technology, and every automation I have built since starts with that discipline.

NextThe same documentation-first pattern applies to the rest of the team's recurring reports. This bridge was simply the largest single one.

How the pipeline is built
02

Cummins · AI prognostics

Recommending against an $8M AI program the data could not support

When the responsible answer was no, and it had to survive leadership pushback

  • AI feasibility
  • Data readiness
  • Model risk
  • Hands-on analysis
  • Executive communication

The situation

A machine learning program aimed to predict failures of an engine component across 50,000 customer vehicles, projecting $8M in warranty savings and an 80% cut in downtime per event. My first mandate was to establish feasibility.

There was no precedent to adapt, the assigned engineer went on leave as the program started, and leadership wanted the answer to be yes. Any no had to survive hard scrutiny.

What I did

  1. Built the domain knowledge. I worked with subject matter experts and sensor vendors to understand 16 sensors and the component health calculations behind them.
  2. Did the analysis hands-on. With data engineers I unified 200+ derived parameters, then ran the exploratory analysis myself in Python on Databricks, overlaying every component's behavior around real failures.
  3. Validated it before it left my desk. I walked the plots and the logic through the subject matter experts and the data scientists separately, and had each group confirm the conclusion from its own expertise, so what reached leadership was a cross-functional finding rather than a program manager's opinion.

The pushback

The reception was hostile in exactly the way I expected. Leadership challenged both the analysis and the tooling, asking whether I had used the most sophisticated data tools available. Because I had done the work hands-on, I could answer in the first person: the finding was about what the data contains, not what the tooling can do, and no algorithm can isolate a component-specific signature that does not exist in the sensor data.

Then I moved the frame to the risk they cared about most. Forcing a model onto this data would generate wrong failure predictions across 50,000 customer vehicles, and each false alarm or missed failure would cost far more in customer trust than the program could ever save. The conversation shifted from whether I had analyzed it correctly to what the responsible decision for the customer was.

Outcome

Claimed entitlement
$8M, 80% downtime cut
Real entitlement on available sensor data
Not reachable
Implementation cost avoided
~$1M
Program time redirected
6 months
Wrong predictions that reached customers
None

Leadership, the business stakeholders, and the customer all accepted the no-go. The method, meaning domain first, then the full parameter space, then a cross-component check and multi-discipline validation, became a template for pressure-testing AI investments before funding them.

Being right is a process, not a moment. The conclusion survived scrutiny because every layer under it, the domain knowledge, the data unification, the hands-on analysis, and the multi-disciplinary validation, had been built to survive scrutiny before the meeting ever happened.

What I would do differentlyPair the no-go with a data collection plan and a scheduled re-review, so the answer reads as not feasible yet, and here is what would change it.

How the analysis was run
03

Amazon · planning organization · 2026

Building AI literacy across a planning team, with no mandate or budget

From one person building with AI to four, self-initiated

  • AI literacy
  • Live demos
  • 1:1 coaching
  • Reusable materials
  • Adoption tracking

The situation

Agentic AI tools were rolling out company-wide, but my team had no structured way to start. A whole team's productivity gain was sitting unclaimed and it was nobody's job. There was no training budget, no formal enablement program yet, and a team that reasonably trusted the workflows it already had. Generic training rarely changes behavior.

What I did

  1. Built depth before teaching. I self-sourced a four to five hour agentic AI workshop and worked through the manual, the examples, and the tutorials end to end.
  2. Showed rather than told. I ran a live demo on a real workflow at the team meeting, rehearsed until it was reliable, because watching an agent do actual work is what converts skeptics.
  3. Coached on their own work. I shared the materials with the team and ran 1:1 sessions mapping the tools onto each person's recurring tasks.
Four designed steps from personal training through a live demo and shared materials to one-to-one coaching, ending in four independent builders Build depth 4-5 hr workshop Show it live whole team Share materials whole team Coach 1:1 2 teammates 4 independent builders capability awareness and reference application adoption
Four designed steps, from my own learning to other people building without me.

Outcome

Teammates now building on their own
4
Live demos on real workflows
1
1:1 coaching sessions
2
Training budget and headcount
$0

Four teammates now apply AI tooling to their own recurring processes without needing me, and I became the team's go-to person on agentic AI.

People adopt when they see their own Monday task automated, not someone else's. I designed awareness, reference, and hands-on application together, and I tracked adoption as a result in its own right.

What I would do differentlyStart enablement in parallel with my own builds instead of after them. Teammates' ideas from the 1:1s would have improved my designs sooner.

04

Amazon · Last Mile carrier mix planning · 2026

An always-on monitoring agent the whole team can question

Network risk surfaced the same day, with no recurring manual effort

  • Agentic workflow
  • Anomaly detection
  • Scheduled automation
  • Conversational delivery

The situation

Planners need to see which carriers and stages are running hot before it turns into a service failure, not after. Reading carrier performance, capacity utilization, and cap-out risk meant manual pulls across several systems, and nothing surfaced anomalies on a schedule, so problems were caught retrospectively. Building an agent needed skills I did not have yet.

What I did

  1. Learned first, built second. I taught myself the company's internal agentic tooling before writing anything.
  2. Built a persistent agent. It runs on a cloud workstation twice a day, combines a trailing and a forward three-day view across three data sources, and applies anomaly detection to flag the three signals planners actually act on: hot stages, cap-out hours, and outlier carriers.
  3. Put the output where people already work. I published through conversational agents into a dedicated team chat channel, so the team can interrogate the read rather than open a static report.

Outcome

Scheduled runs
8:00 AM and 5:00 PM
Data sources combined
3
Window covered
3 days back, 3 ahead
Recurring manual effort
None
Detection of utilization risk
Same day

The whole team queries the channel directly, so network reads no longer depend on one person running pulls.

A report nobody opens changes nothing. Delivering the insight inside the team's existing channel, in a form they can question, is what turned a report into a tool people use.

How the agent is built
05

Cummins · AI prognostics · India market

Delivering an $800K AI pilot after losing the only data scientist

Negotiated a replacement from a team with no capacity, then ramped them myself

  • Delivery to commitment
  • Cross-team negotiation
  • Risk and escalation
  • Stakeholder updates

The situation

The program was built to detect engine failures early enough for planned maintenance instead of roadside breakdowns, targeting a 40% cut in customer downtime. The pilot was the proving ground for making AI prognostics a standard offering in that market.

Halfway through the six-month timeline, the one data scientist building the model left. The data science organization had no spare capacity, my manager was on strategic priorities, and the customer expected a pilot on a committed date.

What I did

  1. Negotiated in the other team's terms. I went straight to the data science manager, without routing through my own, and framed it as business risk rather than headcount: the $800K at stake, the timeline exposure of every unstaffed week, the customer commitment behind the pilot date. He had nothing spare, but he worked the problem with me and offered the one resource that existed, a brand-new hire. I took it on the spot. An inexperienced scientist starting now beat an experienced one starting too late, because the experience gap was a bounded problem I could attack and a calendar gap was not.
  2. Ran the ramp myself. Two intensive weeks on business context, requirements, and the partially built model, so the work continued rather than restarted.
  3. Told the customer early. I re-planned the pilot from two months to two and a half, spent the buffer I had built in, and explained why before they could discover it.

Outcome

Program budget
$800K
Engines with validation data
500
Ramp time for the new data scientist
2 weeks
Overall launch timeline
Met
Customer relationship
Intact

The customer stayed engaged and gave positive early feedback, and the pilot produced the evidence base for AI prognostics in that market.

I asked for help with a risk, not for a headcount, and then made the weak option work. Buffer is a strategic asset to be spent visibly, not a cushion to hide slippage in. Telling the customer two and a half months with reasons earned trust that two months followed by a surprise never could.

How I judged itI deliberately did not measure success as zero slip. After losing the only modeling resource, that metric would have rewarded hiding risk instead of managing it. What I would do differently: engage the data science manager the day the departure was known, and structure the departing scientist's last weeks around knowledge transfer rather than squeezing out final output. I now run a bus-factor check at kickoff, so any critical path with a single named person on it gets a continuity plan before it becomes an emergency.

06

Marelli North America · simulation engineering

Getting four global teams off a decades-old manual workflow

The influence template I have reused ever since

  • Change management
  • Validation against baseline
  • Python automation
  • Influence without authority

The situation

Design of experiments work needed hundreds of simulation models per week to hit customer deadlines. Each one was built by hand in about 30 minutes, the way teams worldwide had done it for decades.

I was a junior engineer with no mandate. The tool's Python interface was brand new and unproven, and a wrong model could mean a wrong part shipped to a customer.

What I did

  1. Found the real leverage point. About 80% of parameters stayed constant between runs, so I built one parameterized model instead of speeding up manual rebuilding.
  2. Built a credibility ladder. I designed each step to produce proof I could show a skeptic later, not just working code: a sample script and documentation from the software's support team, then a simple model run end to end, then a production-grade one carrying the real geometry, boundary conditions, and parameter interactions. Every rung was verified against the manually built equivalent before I advanced, and the manual path stayed alive as a fallback until the automated one had earned its place.
  3. Drove adoption team by team. The teams were hesitant to leave a method they had trusted for decades, which I expected. Rather than argue, I applied the automation to their models, their design nuances, their validation criteria. Each region's models had local quirks the script had to absorb, so this was real engineering work done with them, which improved the tool and transferred ownership at the same time. A first failure on someone's own model would have killed adoption permanently.
Thirty minutes per model manually against five minutes scripted, and the split showing eighty percent of parameters unchanged between runs Minutes per model Manual 30 Scripted 5 What changed between runs ~80% of parameters stayed constant ~20% changed
The insight that unlocked it: the answer was one parameterized model, not faster manual rebuilding.

Outcome

Time per model iteration
30 min to 5 min
Faster per model
83%
Total modeling time removed
~80%
Global teams standardized on it
4
Accuracy against the manual baseline
Preserved

One by one, all four global teams adopted the parameterized approach on their own use cases. It became the standard way to build these models across regions for the first time, and new work started from the script instead of a blank manual model.

I never asked the teams to trust me. I asked them to watch their own model build in five minutes instead of thirty, with results matching their baseline. People do not adopt your solution, they adopt their results.

What I would do differentlyPackage the documentation and worked examples from day one, so adoption scales without me in the room. I drove it through personal hands-on sessions, which worked but made me the bottleneck. I also treated it as a side project longer than I should have; raising it with management earlier would have resourced the rollout properly.

How I deliver an AI use case

The same sequence runs through every program above, from requirement to realized value. Each step links to where it happened.

  1. Start from the business problem

    Turn the pain into a requirement and a success measure before choosing a tool. The weekly bridge was six manual steps and 2 to 3 hours; success meant matching the bridges leadership already trusted.

    Variance narratives
  2. Check whether the data can carry it

    Before any model, test whether the data holds a usable signal. Sixteen sensors and 200+ parameters said it did not, and the right call was to stop.

    AI feasibility
  3. Prototype against the current process

    Build the first version myself and benchmark it against the trusted baseline: past reports, manually built models, the outputs people already sign off on.

    Variance narrativesGlobal adoption
  4. Keep a human in the loop until it earns trust

    Review outputs through an initial period and keep the old path alive as a fallback. Unattended runs come after the track record, not before.

    Variance narrativesGlobal adoption
  5. Deploy where people already work

    The Monday email, the team chat channel, the existing template. Output nobody opens changes nothing.

    Monitoring agentVariance narratives
  6. Build capability, not dependency

    A live demo on real work, shared materials, then 1:1 coaching on each person's own task. The measure that matters is how many people build without me.

    AI literacy
  7. Report realized value honestly

    State what was captured against what was promised, including when the answer is no, and what would change it.

    AI feasibilityAI pilot

How it's built

The technical detail behind three of the programs above: what each system does, how outputs were checked before anyone relied on them, and the design choices that mattered. Internal system names are generalized.

Build 01 · Amazon

Variance narrative pipeline

Read the case study

An unattended agent that runs the weekly carrier mix and rate bridge end to end: queries the data, builds the tables, writes the variance narrative, checks its own work, and emails the finished package every Monday.

Data warehouse feeds an agent that is guided by a workflow specification and a set of historical bridges. The agent produces a Word report, an Excel workbook and a variance narrative, which pass human review during an initial period before an unattended Monday email. Review corrections feed back into the agent. Workflow spec sources, queries Past bridges format and tone Data warehouse SQL queries Agent run 1 run encoded queries 2 fill output templates 3 draft variance narrative 4 run validation checks Word report Excel workbook Variance narrative Human review initial period, then unattended Monday email daily in peak corrections fed back until outputs matched past bridges
Past bridges do two jobs: they set the format and tone the agent writes in, and they are the standard early outputs were graded against.

Components

Inputs
Queries against the internal data warehouse: the same logic the manual process ran, written down first.
Specification
The documented workflow: data sources, query logic, and the output table templates.
Reference set
Historical bridges, used as worked examples for structure and tone, and as the acceptance standard.
Outputs
A formatted Word report, an Excel workbook, and the written variance narrative, sent as one email package.
Cadence
Unattended every Monday morning with no manual trigger; run daily during a major sales event against an 8:00 AM deadline.

Validation and controls

  • Named the failure mode first. A wrong number reaching a leadership document without anyone noticing.
  • Validation built into the workflow, not bolted on after.
  • Outputs anchored to past bridges for structure, tone, and interpretation.
  • Human review through an initial period; unattended only once the record held. The daily peak cadence was the hardest live test.

Design choices

  • Documentation is the specification. An agent can only be as good as the written process it executes.
  • One reference set, two jobs. The same examples shape the output and grade it.

Build 02 · Amazon

Network monitoring agent

Read the case study

A persistent agent that reads the delivery network twice a day, flags the signals planners act on, and answers follow-up questions in the team's chat channel.

A schedule triggers the agent at 8 AM and 5 PM. It ingests three internal data sources, builds a six-day view spanning three days back and three days ahead, runs anomaly detection for hot stages, cap-out hours and outlier carriers, and posts to a team chat channel where teammates ask follow-up questions. Schedule 8:00 AM 5:00 PM Ingest 3 internal sources carrier data 6-day view 3 days back 3 days ahead Anomaly detection hot stages cap-out hours outlier carriers Team chat channel posted by conversational agents Teammates ask follow-up questions No manual trigger. No recurring manual effort.
The forward three-day window is the point: it turns monitoring from a record of what went wrong into a warning about what is about to.

Components

Runtime
A persistent agent on a cloud workstation, started on schedule twice a day.
Inputs
Carrier performance and capacity data from three internal data sources.
Window
Trailing three-day performance and utilization, combined with forward three-day utilization.
Detection
Anomaly detection on three signals: high-utilization stages, cap-out hours, and outlier carrier performance.
Delivery
Conversational agents in a dedicated team channel, so the read can be questioned rather than just opened.

Design choices

  • Forward-looking, not just retrospective. Risk is caught the same day, before it becomes a service failure.
  • Conversational delivery. A report nobody opens changes nothing; a channel people already use, which answers back, does.
  • Learn the platform before building on it. I taught myself the internal agentic tooling first, so the design was not limited by what I did not know.

Build 03 · Cummins

AI feasibility analysis

Read the case study

Not a model, but the test a model has to pass before anyone funds it: does the data contain a signal specific enough to predict this failure? Run hands-on in Python on Databricks.

Analysis path from sixteen sensors to two hundred parameters to a cross-component comparison, ending in a no-go decision 16 sensors understood 200+ parameters unified Compare every component at failure Python on Databricks No unique signature the whole system moved together Decision: stop Validated independently by subject matter experts and data scientists before it reached leadership.
The target component alone looked promising. Only the comparison across the whole system showed the signal was not specific to it.

Method

  1. Domain model first. Component health calculations and normal operating thresholds with subject matter experts, plus sensor vendor catalogues for what each sensor could physically resolve.
  2. The full parameter space. 200+ derived parameters unified into one format with data engineering, rather than a convenient subset.
  3. Event-window analysis. How every parameter moved in the time windows around actual component failures.
  4. Isolation test. Overlaying every component's behavior at failure, to see whether any pattern belonged to the target alone.
  5. Independent validation. Subject matter experts and data scientists each confirmed the finding from their own expertise before leadership saw it.

Viability criteria

  • Consistent data to learn from.
  • Sufficient parameter coverage of the system.
  • A repeatable pre-failure signature for the target component. This one failed.

Design choice

  • Test whether a signal is specific, not just present. A model trained on system-wide noise would have produced wrong predictions on customer vehicles.

What I build with

  • Python
  • SQL
  • Prompt engineering
  • Agent orchestration
  • AWS Bedrock
  • Athena
  • S3
  • QuickSight
  • Amazon Quick
  • Azure Databricks
  • Power BI
  • Advanced Excel
  • ChatGPT Enterprise
  • Copilot

Background

Three industries, one pattern: inherit an ambiguous problem, root-cause it with data, leave behind a capability the team keeps.

  1. 2025 to present
    Sr. PM, Planning and Strategy · AmazonLast Mile carrier mix planning for the Canada delivery network. Built the variance narrative pipeline and the monitoring agent above, and led AI enablement for the planning team.
  2. 2024 to 2025
    Sr. Manager, AI Enablement and Automation · CumminsLed a team of four technical program managers delivering AI and advanced analytics programs across five concurrent programs.
  3. 2022 to 2024
    Sr. Technical Program Manager, AI Enablement and Automation · CumminsAI prognostics and predictive maintenance programs, including the feasibility review and the India pilot above.
  4. 2019 to 2022
    Sr. Program Manager, NPI · Marelli North AmericaProgram lead for automotive hardware programs at a Tier-1 supplier, for a Top 3 US automaker.
  5. 2017 to 2018
    Hardware Product Engineer · Marelli North AmericaSimulation and test automation in Python, including the global adoption case above.
  6. 2016
    Vehicle Engineering Intern · TeslaMultiphysics models for Model 3 thermal components.

Education

  • 2020 to 2023
    MS, Data Science · Indiana University Bloomington
  • 2015 to 2017
    MS, Mechanical Engineering · Michigan Technological University
  • 2009 to 2013
    BS, Mechanical Engineering · College of Engineering Pune

AI and automation

  • Generative AI and prompt engineering
  • Agentic workflows and agent orchestration
  • Narrative generation and AI-assisted analysis
  • Anomaly detection monitoring
  • AI feasibility and data readiness
  • Human-in-the-loop validation

Planning and analytics

  • Forecasting and forecast accuracy
  • Variance analysis and management reporting
  • SQL, Python, Databricks
  • Power BI and advanced Excel
  • Scenario and sensitivity analysis

Delivery and enablement

  • Use case scoping and success measures
  • Delivery with technology teams
  • AI literacy: demos, materials, 1:1 coaching
  • Change management and adoption tracking
  • Executive communication and escalation