Selected work

Problem, system, number, lesson.

Most companies stay unnamed. Avon Seals is on the record. The rest are EdTech B2B and prior roles — outcomes sanitized, methods not.

“If the tool is not live before the room, it is theatre.”

Avon Seals · Chennai · 01

Live

assessment app

Baseline before the room

Management Development Assessments for a leadership program I delivered. The artifact is an interactive app — still live — not a slide deck after the fact.

Open the live app

01 · Problem
A leadership program without a baseline is a workshop, not development. Managers would walk into orientation with no shared picture of how they delegate, plan, or how the team debates and commits.
02 · System
An assessment app taken before orientation: Baseline Managerial Assessment (delegation, prioritization, planning, communication, decision-making) and a team-health check (trust, debate, commitment, accountability, results). Screenshot the results; bring them into the room.
03 · Outcome
The product shipped and still runs. Baseline first, then the session — not a score I invented after the fact.
04 · Lesson
Enablement is a system. If the tool is not live before the room, it is theatre.

EdTech B2B · 02

~$125k

orphan ACV

Renewals no rep could see

A Slack nudger fused live CRM and call notes so renewals without an owner stopped being invisible.

01 · Problem
Renewals sat in HubSpot with no active owner. Reps could not see them on a list. Risk showed up as a surprise, not a Tuesday-morning action.
02 · System
HubSpot plus Gong, written into plain-English Slack blocks: why this account is here, what to do next. Read-only. Named CS/BD sponsor.
03 · Number
One run put 41 renewals in front of owners, including ~$125k ACV with no owner.
04 · Lesson
Automation earns trust when it returns dollars and a next step — not another chart.

EdTech B2B · 03

17/21

reps matched baseline

Hygiene as a metric, not a meeting

Commission-tier scoring BD already believed, made reproducible off live HubSpot.

01 · Problem
Close date, next step, and activity were argued in the room, then forgotten by Friday. Hygiene was a review, not a number.
02 · System
Nightly Green / Yellow / Red decelerators leadership had already agreed. Criteria in the open. No black-box model. 460 deals on first run.
03 · Number
17 of 21 reps matched the hand-built baseline the first night.
04 · Lesson
Automate the score people already argue about. Then get out of the way.

EdTech B2B · 04

14/15

LLM-as-judge

The AI SDR had to pass a test first

WhatsApp outreach with deterministic guardrails and a judge harness before anyone talked to a customer.

01 · Problem
A non-English AI SDR is a brand risk on honorifics, opt-out, and price leak. “We’ll watch it in production” is not a gate.
02 · System
FastAPI agents, lint for the failure cases that matter, 15-scenario LLM-as-judge. Draft mode until the grade cleared.
03 · Number
14 / 15 on the harness before go-live.
04 · Lesson
Ship the eval suite before the agent. Receipts beat vibes.

EdTech B2B · 05

256/256

checks vs the old board

Board ARR you can reproduce

Replaced the Excel and Looker argument with Snowflake truth and a written Metrics Book.

01 · Problem
Finance and sales could not agree on ARR / NRR / GRR. The number of the week depended on who built the tab.
02 · System
Snowflake / Streamlit with SQL lineage, plus a Metrics Book so leadership numbers are documented, not political.
03 · Number
256 / 256 checks against the prior board tiles.
04 · Lesson
Governance is a document plus a pipeline, not a vibe.

Prior role · telecom · 06

57%

cluster cost down

Same judgment. Different surface.

The craft I now use on CRM and agents: utilization criteria, written forensics, no one-off scripts.

01 · Problem
A production cluster was expensive, and the diagnosis lived in someone’s notebook. Leadership could not replay the argument.
02 · System
Utilization criteria and reproducible cost diagnostics — the same standard I now apply to hygiene scores and renewal risk.
03 · Number
57% reduction in cloud cost on the cluster we measured.
04 · Lesson
I used to optimize models. Now I optimize whether a renewal has an owner. The standard did not change.