Latest ·
Thoughts vs thinking: the distinction that quietened my mind
Thoughts arrive on their own. Thinking is what you do with them. Two writers made this click for me, and then I found the Bhagavad Gita had mapped it out.
A career in three acts · told out of order
Kalpak Kale
Operations Manager at Latent (Alpha OBS LLP). Twelve years, from the edit suite to the operations desk.
Dispatches
Field notes on putting AI to work · evaluations, agent workflows, shipping without a dev team · by Kalpak Kale
Act I
Present day
October 2023 onward.
Scene 01
The operations desk
I keep a sixty-person services firm delivering, and build what nobody else has time to build.
- 15client projects in flight at any time
- 309items tracked over six months
- 26office tech fixes for 12 colleagues
The full story
As Operations Manager I sit between the project managers who hold the large accounts and everything else that has to happen for a sixty to seventy person services firm to deliver. I lead AI product development from scoping and architecture through deployment and post-release evaluation, and I act as delivery manager on the smaller client projects across BFSI, telehealth, e-commerce and government work.
- Delivery manager on smaller client projects alongside the project managers: around 309 tracked items over six months.
- Internal auditor for the CMMI Level 3 implementation: built and populated the Process Asset Library, maintain the training repository, and completed audit interviews for 15+ people across sales, HR, staffing, project management and development, with the readiness assessment presented to leadership.
- Own hosting and deployment for assigned projects; manage the company's Workspace, GCP, DNS and hosting accounts; support GeM and government tenders.
- The internal fixer for office technology: around 26 requests resolved across 12 colleagues in six months.
Dispatches · recent
Recent dispatches
What worked, what broke, and the numbers behind it. All dispatches · Daily news updates · RSS
Thoughts vs thinking: the distinction that quietened my mind
Thoughts arrive on their own. Thinking is what you do with them. Two writers made this click for me, and then I found the Bhagavad Gita had mapped it out.
Every message is a seed: designing an intent-first RAG chatbot
Design pattern for trustworthy chatbots: treat each message as a seed, context as soil, retrieval as nourishment, turn values into testable response rules.
The 21-case test that decides whether a RAG chatbot ships
A document chatbot for a regulated financial client went from a proof of concept to client beta. What got it there was not the model. It was a weighted evaluation of 21 cases, nine rounds of structured QA, and a daily accuracy report the client ran themselves.
Giving 19 developers AI access through one proxy, and what actually changed
Rolling AI coding tools out to a team is not just licensing. It is a gateway, spending policy, evaluation habit and working rules for a 60-person firm.
A brief goes in, software comes out. Shipping client products without a dev team assigned.
An operations manager with an agentic coding tool can ship a statement of work to production using a fixed process for payments and security.
Act II
Two decades earlier
Shegaon, 2004. A four-year degree begins.
Scene 02
Trained on the job first
Before the degree was finished, I had already kept a bank's systems running and sold door to door.
- 24years old, supporting a bank's head office on site
- 2trades learned before graduating: sales and systems support
- 1habit that never left: finish what you start
The full story
Most people leave engineering college with a degree and no work history. I left with both, in the other order. While the degree was still in progress I sold door to door for an NGO, a month of hearing no from strangers and learning to ask anyway, and then spent six months as Resident Engineer for Gayatri Infotech at the Bank of Maharashtra's central office in Shivajinagar, Pune: desktops, network, users who needed things fixed today.
The degree itself, at Shri Sant Gajanan Maharaj College of Engineering, ran from 2004 to 2012 instead of the planned four years. I mention it because it is where two things I rely on every day were formed. I finish what I start, however long it takes. And I learned operations by doing it inside a bank before I read about it in a textbook. When I sit at the operations desk now, that is still the room I am in.
- Resident Engineer, Gayatri Infotech, on site at the Bank of Maharashtra central office, Pune · , six months
- Door-to-door fundraiser, NGO · one month
- B.E., Information Technology, Shri Sant Gajanan Maharaj College of Engineering, Shegaon · to
Scene 03
The edit suite
It started with film. Cut it, watch it end to end, then ship it.
- Video Editor · 2013 to 2014
- Film and TV Editing certificate · 2012
- First job after the degree finally closed
The full story
I started in film. A Certificate in Film and TV Editing from Digital Academy in Mumbai led to a video editor's chair at Indian Magic Eye, cutting corporate films from footage handover to final delivery. Pacing, continuity, quality control before anything went to a client.
Editing taught the habit that has followed me everywhere since: the work is not done until someone has watched it end to end and it holds together.
- Video Editor, Indian Magic Eye Pvt. Ltd. · to
- Certificate in Film and TV Editing, Digital Academy, Mumbai ·
Scene 04
Pixels for forty languages
Localised e-learning imagery for learners in dozens of languages.
- Graphic Designer and Integrator, Lionbridge · 2015 to 2016
The full story
At Lionbridge I localised e-learning imagery by language and region, then reassembled every version in the authoring tool and checked it for layout and language accuracy. Detail work at volume, where a single wrong glyph ships to thousands of learners.
- Graphic Designer and Integrator, Lionbridge India · to
Scene 05
Into the server room
Three months at Prime Focus. Then on site at Disney, keeping a media company's systems up.
- Infrastructure Engineer, Prime Focus Technologies · 2016 to 2017
- Deployed to Disney's Mumbai office after three months
- On-site client support, as at the bank, now for a media company that cannot afford downtime
The full story
Prime Focus Technologies moved me from pixels to packets. The first three months were at their Andheri MIDC office. Then they deployed me as Infrastructure Engineer to Disney's Mumbai office, also in Andheri MIDC, to maintain the Prime Focus systems installed at the Disney site.
The role was the same shape as my Bank of Maharashtra posting: my employer's systems, installed in the client's building, kept running by me on site. The difference was the cost of failure. A media company cannot afford downtime, and people needed things fixed today. Preventive maintenance, planned upgrades, Linux troubleshooting over remote access, and walking people through hardware fixes step by step.
This is where hosting, DNS and "why is it down" stopped being someone else's problem, and where I learned to represent one company inside another. Most of my work since has had that shape.
- Infrastructure Engineer, Prime Focus Technologies · to · three months at the Andheri MIDC office, then on site at Disney's Mumbai office
Scene 06
Running projects
A consultancy first, then from QA analyst to Head of Projects in three years.
- Alpha OBS LLP, now Latent · 2020 to 2023
- Web, mobile and e-commerce builds
- Before that: inventory-module rollout at Suhrud Consultants, 2017 to 2018
The full story
After Prime Focus I moved into project management at Suhrud Consultants in Navi Mumbai, then edited films for clients of my own through 2019 and the first half of 2020. In June 2020 I joined Alpha OBS LLP, the company now known as Latent, as a quality analyst and supervisor, moved into project management, and then took over the projects function as Head of Projects. Web, mobile and e-commerce builds for concurrent clients; deadlines set, progress tracked, issues unblocked during execution rather than at review time, and the client conversations when scope or timelines had to change.
At Suhrud the job was an inventory-module rollout, assessing stock-management operations and finding the ERP data-capture gaps that would have produced wrong stock positions. The wedding films paid for the gap year before Latent.
- Quality analyst, then Project Manager and Head of Projects, Latent (previously Alpha OBS LLP) · to
- Project Manager, Suhrud Consultants, Navi Mumbai · to
- Freelance Video Editor · to
Scene 07
Back to school, repeatedly
Studied operations while doing it.
- IIM Lucknow · Executive Programme in IT Management · 2023
- IIM Mumbai · Generative AI for business decisions · 2026
- MBA and PG Diploma in Operations Management
- B.E. Information Technology · 2012
The full story
Operations is a discipline, so I studied it: a postgraduate diploma and an MBA in operations management, business analytics, and then the Executive Programme in IT Management at IIM Lucknow while working full time. In 2026 I added IIM Mumbai's programme on generative AI for business decision-making, because the tools I was already using deserved a proper theoretical footing.
- Generative AI for Data-Driven Business Decision-Making, Indian Institute of Management, Mumbai ·
- Executive Programme in IT Management, Indian Institute of Management, Lucknow · to
- Certification in Business Analytics, EduPristine, Mumbai ·
- Post Graduate Diploma in Operations Management, MITSDE, Pune ·
- MBA, Operations Management, Jaipur National University (distance learning)
Certifications
- WhatsApp Business Platform for Developers, Meta ·
- Create Interactive Dashboards with Streamlit and Python ·
- Claude Certified Architect to Foundations · in progress
Act III
Convergence
2025. The editor, the engineer and the manager meet.
Scene 08
A brief becomes software
A brief goes in. Working software comes out. No development team assigned.
- 19developers given AI access
- 38tools evaluated, each with a written verdict
- 12engineers in the Claude Architect programme
The full story
The biggest change in how I work: I now specify, build and ship working systems with agentic coding tools, taking a brief to deployed software without a development team assigned. Not demos. Production systems with payments, encryption, QA cycles and clients using them daily.
- Set up AI access for around 19 developers through a self-hosted LiteLLM proxy, standardised agentic workflows, and enrolled 12 engineers in the Claude Certified Architect to Foundations programme.
- Build custom MCP servers connecting the internal CRM, memory store, issue tracker, hosting provider and Google Drive to daily tooling, and author reusable Claude skills.
- Ran roughly 38 structured tool evaluations, each closing with a written recommendation, before anything reached a client.
- Autonomous agents with human approval gates now handle content, sourcing and reporting work that used to be manual.
Selected work
Four problems. Four systems.
Each one is live or in client hands. Names left out on purpose.
01
Travel entry platform
A traveller needs a different government form for every destination. Nothing may be recorded until the money clears.
6destination workflows live in production
- 1Traveller picks a destination
- 2Rules engine shows only the fields that apply
- 3Payment clears first
- 4Application recorded once, reconciled by cron
- 5Encrypted PII, signed document links
Details
Problem. Travellers need a different government form for every destination, each with its own fields, fees and eligibility rules. Originally scoped for tourist visas, the product now runs as an arrival-card system.
My part. Owned it end to end. A payment-first submission model so an application is only recorded once payment clears, made idempotent across the webhook, the return page and a cron reconciler. PII encrypted at rest with AES-256-GCM, documents behind short-lived signed URLs, uploads scanned by magic byte. Ran the QA cycle: test, file, fix, deploy, verify, close.
Outcome. Live in production with multiple destination workflows enabled.
02
AI document chatbot
A wealth-management team needs accurate answers from thousands of documents without stepping outside RBI and SEBI lines.
9structured QA rounds before client beta
- 1Documents ingested, slides read by vision model
- 2Indexed in Milvus
- 3Question retrieves the right passages
- 4Answer with citations, inside compliance scope
- 521-case evaluation before every release
Details
Problem. A BFSI and AIF client needed staff to get accurate answers from a large document corpus without breaching RBI and SEBI compliance boundaries.
My part. Proof of concept to demonstrable prototype: rebuilt a React, Flask and FAISS script onto a containerised FastAPI, Celery, Redis, Milvus and PostgreSQL stack behind Traefik with GHCR-based CI. Designed a 21-case weighted evaluation framework covering intent, retrieval and semantic drift, ran nine structured QA rounds, held the client feedback calls and made the beta release call.
Outcome. In client beta with daily accuracy reporting, and a vision-based ingestion path that lifted slide-deck text extraction well above standard OCR.
03
Order management system
A jewellery manufacturer takes orders as handwritten slips and voice notes. Vendors must never see customer identity or price.
180tracked issues delivered in production
- 1Slip photographed, voice note recorded
- 2Claude reads the slip, Whisper transcribes the note
- 3Order created and tracked
- 4Customer updated on WhatsApp
- 5Vendor receives design spec only
Details
Problem. Orders arrived as handwritten slips and voice notes, and vendors needed design specifications without ever seeing customer identity or pricing.
My part. PHP 8 and MySQL REST backend with a React 19 front end; Claude OCR reading order slips and Whisper transcribing voice notes; a WhatsApp Business webhook for customer and vendor messaging with vendor messages scoped to design specifications only.
Outcome. Live in production with around 180 tracked issues delivered, including the Meta app review.
04
Autonomous agent workflows
Content, prospecting and reporting were manual, repetitive and easy to let slip.
60prospects and candidates sourced per day
- 1Agent audits the site and Search Console
- 2Researches keywords, drafts the article
- 3Emails it for approval
- 4Applies the reply, publishes, resubmits
- 5Scheduled agents feed prospects and candidates daily
Details
Problem. Content, sourcing and reporting work was manual, repetitive and easy to let slip.
My part. A content pipeline that audits the site, pulls Search Console and analytics issues, researches keywords, drafts an article, emails it for approval, applies the reply revisions, publishes and resubmits to Search Console. Scheduled agents add 50 prospect companies and 10 sourced candidates daily; a weekly cron refreshes a KPI dashboard across three properties.
Outcome. Recurring work runs unattended with a human approval gate where it matters.
- 40+client prototypes in six months
- 3production web properties
- 180tracked issues delivered on one production system
- 12years across media, infrastructure and IT services
Also built and shipped
- More than 40 client prototypes in six months across government, BFSI, healthcare and consumer sectors: secure messaging for state police and defence commands, a deepfake voice-detection demo for a national fraud-reporting service, scheme portals, analyst and credit tools, clinical dashboards. Several went into investor pitches and converted to paid work. Wrote the SOP and trained a colleague to run the process.
- Internal systems: an asset management system on TanStack Start and MySQL replacing the company asset register; a FastAPI payroll analyser with OCR parsing and increment simulation; an Electron, Node.js and MongoDB employee monitoring system; Apps Script payroll and subscription trackers.
- Three production web properties (an in-house product, a client site and the company publication): JSON-LD schema, canonical and redirect handling, CSP and xmlrpc hardening, Core Web Vitals, GA4 and Search Console via service accounts, rsync deploys with rollback packages. Shipped a Manifest V3 Chrome extension to the Web Store and root-caused a live site-down incident to a disabled plugin.
Toolkit
Picked up in this order
Fifteen years, one tool at a time. Each bar is when a capability entered the kit and whether it is still in daily use.
Operations and delivery
AI and automation
Technical working knowledge
Web operations
Earlier crafts
Epilogue
Off the clock
Scene 09
Higher ground
Cadet, trekker, occasional editor.
- NCC "A" Certificate
- National Himalayan Trekking Expedition · 2002
- Marathi, English, Hindi
The full story
Before any of this I was a cadet. An NCC "A" Certificate with 11 Maharashtra Battalion in Akola, a National Himalayan Trekking Expedition in 2002, and the National Standard Examination in Physics in 2004. I still travel and trek when I can, still cut the occasional film, and spend a suspicious amount of free time evaluating new AI tooling.
Languages
- Marathi (native)
- English (fluent)
- Hindi (fluent)
Interests
- Travelling and trekking expeditions
- Film and video editing
- Evaluating new AI tooling and turning it into automations
At a glance
Public details
- Location
- Pune / Akola, Maharashtra, India
- Company
- Latent (Alpha OBS LLP)
- Role
- Operations Manager
- linkedin.com/in/kalpakkale
- Medium
- medium.com/@kalpakkale
- Website
- kalpakkale.com
Roll credits
Contact
Write to kalpak@kalpakkale.com, connect on LinkedIn, or use the form. It lands in my inbox and I reply to the address you give.