01 · The question
The people analytics team you were going to hire. Live in 90 days, inside your own cloud.
We build and run your people analytics function, so you keep more of the people you wanted to keep, spend less on people operations, and turn engagement into action. Built by people who led these teams inside Indian enterprises.
Don’t take our word for it. Ask it something.
Pick a question a CHRO asked us this month
Three stores have run above 1.5× the West zone early-exit rate for two months: Andheri East hub 23.8%, HSR Layout 19.2%, Gomti Nagar 17.9%. Zone average 11.2%. The pattern is already in the last 60 days; no model needed for this one.
Grounded in attrition_voluntary_rate v1 · hrms as of 07 Nov 2026 · teams under 10 withheld
Answers come from the semantic layer, not from the model. Same question, same number, every time.
Three stores have run above 1.5× the West zone early-exit rate for two months: Andheri East hub 23.8%, HSR Layout 19.2%, Gomti Nagar 17.9%. Zone average 11.2%. The pattern is already in the last 60 days; no model needed for this one.
Grounded in attrition_voluntary_rate v1 · hrms as of 07 Nov 2026 · teams under 10 withheld
₹2.4 Cr in Q2 FY27: 312 regrettable exits × ₹77,000 average replacement cost (recruiting, 21 vacant days, 30-day ramp). Finance signed off the cost model on 14 Aug 2026, so this is their number too.
Grounded in cost_of_attrition v2 · exits from hrms, costs from finance model fin-2026-08 · as of 07 Nov 2026
61% in October, against the policy target of 84%. Lowest at Hinjewadi hub: 2 of 9. Margdarshak has a nudge drafted for the hub manager; it is waiting for HRBP approval.
Grounded in onboarding_checkin_day7 v1 · hrms + onboarding app · as of 07 Nov 2026
47 of 1,212 managers. 31 of them run dark stores in the South zone, where the average span is 26. ISO 30414 reports span by level; here it is by role, because that is how your org runs.
Grounded in span_of_control v1 · hrms org tree · as of 07 Nov 2026
Runs today in
- Telecom
- Quick commerce
- Fintech
- Internet services
- Healthcare
02 · What it costs today
HR is seen as a cost because its value is never measured
The data to fix this already sits in your systems. What is missing is a function that reads it, acts on it and measures the result.
| Outcome | How it is done today | With Dattamsha |
|---|---|---|
| Lower regretted attrition | Attrition is found out at the exit interview and reported a quarter late; HR and Finance quote different numbers. | Teams at risk are flagged nightly; the manager gets one approved nudge the same week. |
| Lower cost of people operations | Reports built by hand; the same policy questions answered again; the comp cycle on spreadsheets and email. | One view per manager and HRBP; answers that quote the clause; comp data to letters in one tool. |
| Higher engagement, with action | A survey ends in a results deck. A dip seen in March is acted on in June. | Short pulses built on behavioural science; each result becomes one action for the manager. |
| Lower compliance risk | Your DPO or privacy lead asks for a data map; the answer is a spreadsheet from 2024. | Consent, retention and erasure recorded per employee; the evidence pack is one export. |
The arithmetic
What one point of attrition costs a 2,000-person company
2,000 employees × 1 point = 20 exits a year
20 exits × 6 lakh rupees to hire and ramp each = 1.2 crore rupees a year
Illustrative assumptions for salaried roles; your own figure comes from your data by Day 90.
03 · The night before
“Why are riders in Pune leaving in their first 30 days?”
That question lands on Monday. Right now the answer arrives on Thursday, in a deck, with numbers Finance disputes.
Here, the analyst agent noticed on Sunday night, investigated in five steps you can read, and drafted the one action. By Monday morning it is waiting for your HRBP to approve.
Press play to watch it think.
Early exits, Pune riders: 23% in Oct vs 13% in Sep
- 23:12
Signal from Gargi
First-30-day exits among riders, Pune zone: 14 of 61 October joiners, vs 9 of 70 in September.
- 23:14
Checked sibling zones
Mumbai 6 of 58, Nagpur 3 of 22. Pune is the outlier, not the season.
- 23:19
Split by hub
Hinjewadi hub accounts for 9 of the 14. One shift lead changed on 22 Sep.
- 23:23
Cross-checked onboarding
7 of the 9 had no Day-7 check-in logged. Policy says one by Day 7.
- 23:25
Drafted a nudge
To the Hinjewadi hub manager, via Margdarshak. Waiting for HRBP approval.
Investigating…
04 · What it takes
That Sunday night needs three things most HR teams don’t have: a data layer, people to run it, and proof it worked
Hiring that team takes a year. We bring it as a function, with the platform underneath, and run it with you.
The data layer
Sangam™ संगम
One governed employee record across HRMS, ATS, LMS and surveys. 58 metrics to ISO 30414 that do not change between runs, because they are precalculated in a semantic layer, not in a spreadsheet.
Every door installs Sangam underneath.
The team
Gargi + Birbal गार्गी · बीरबल
Our analysts and our agents run your weekly cadence: the Monday brief, the signals nobody asked about, the answers to the questions leaders actually ask.
The proof
Phala™ फल
Every nudge is measured against the teams that were not nudged. Every quarter closes with a value statement in rupees, with how each number was measured.
05 · AI, in three verbs
AI that finds the signal, acts on it and answers anyone
The model never sees an employee record. Every number comes from the semantic layer, not from the model. A person with the right role approves every action.
Finds
Birbal picks up a signal nobody asked about, investigates in up to twelve shown steps and keeps a case file you can reopen.
Birbal बीरबल
Monday brief · West zone
1. Pune early exits 23% vs 13% · case open
2. Overtime over 60 h at 3 dark stores · watching
3. Survey response 41% in Nagpur vs 68% · nudge sent
Acts
Margdarshak™ turns the approved finding into one timely nudge to the right manager, measured against those not nudged.
To the Hinjewadi hub manager
7 of your 9 October joiners have had no Day-7 check-in. A 15-minute call this week lowers early-exit risk.
Answers
Sutradhar™ answers a new joiner’s policy question and quotes the clause. Gargi answers a zonal head from the governed metrics, as you saw above.
Sutradhar सूत्रधार
New joiner, on WhatsApp: “Can I carry forward unused leave?”
Yes, up to 30 days of earned leave carry forward each April. Casual leave lapses on 31 March. Leave policy v4, section 5.2.
Quoted from Leave Policy v4 · 19 in 20 right on 304 real questions at a fintech client
Screens show illustrative data.
06 · How it’s built
One metric, shown all the way down, because the number has to survive a Finance review
Nobody in HR software shows you their definitions. Here is ours for voluntary attrition, as it runs tonight.
-- metrics/attrition_voluntary_rate.sql · v1 · ISO 30414 turnover rate
-- Runs nightly in Airflow. Versioned. Same inputs, same number.
with exits as (
select employee_id, exit_date
from mart.fct_employee_events
where event_type = 'exit'
and exit_reason in ('resignation', 'absconded') -- voluntary only
and exit_date between :period_start and :period_end
),
headcount as (
select avg(active_headcount) as avg_headcount
from mart.fct_daily_headcount
where day between :period_start and :period_end
),
scoped as (
select e.*, o.team_id, o.zone_id
from exits e join mart.dim_org o using (employee_id)
)
select
team_id,
case when avg_headcount < 10 then null -- withheld under DPDP policy
else count(*)::numeric / avg_headcount end as attrition_voluntary_rate,
count(*) as exits,
avg_headcount
from scoped cross join headcount
group by team_id, avg_headcount;
The rule, in plain language
- What counts
- Resignations and absconding. Not terminations, contract ends, deaths or internal transfers.
- Denominator
- Average active headcount over the period, not opening headcount. Finance signs this off once.
- Withheld
- Any team under 10 people shows as withheld, never as a number.
- Versioned
- v1 since 14 Aug 2026. A definition change is a new version, so last quarter’s number never moves.
- Open source
- Postgres, Airflow, dbt Core and Soda checks. The code sits in your repository.
Inside your walls
Deployed in your own cloud, served from India. Anvaya™ runs inside your account, in your region. We never host your data.
DPDP-compliant by design. Per-tenant isolation, row-level access, a consent ledger in Sangam, and identity checks through Setu so Aadhaar and PAN never sit in the system.
What we refuse to build
- A dashboard nobody opens. We send the brief where people already read.
- A model that sees an employee record. It sees masked sentences, never the row.
- A saving claimed without a holdout. No comparison group, no number.
- A number that changes between runs. Definitions are versioned and precalculated.
- Lock-in. Open source underneath, deployed in your cloud, the code is yours.
07 · The first 90 days
From the first Monday to the first read-out in rupees
What we do, and the little you give. By Day 90 the Sunday-night case file above is routine.
- Weeks 1–2ConnectSangam joins your HRMS, ATS and surveys. You give us read access and one data owner.
- Weeks 3–5DefineMetric definitions signed off with HR and Finance. You give us two hours a week.
- Weeks 6–8Switch onGargi, Birbal and Sutradhar go live in your cloud. You name the approvers.
- Weeks 9–12ActFirst nudges go out with a holdout. Your HRBPs approve; we run the cadence.
- Day 90Read outPhala’s first value statement, in rupees, with what was and was not measured.
08 · Where it runs today
Five enterprises: the problem, the change, the result
09 · The platform, briefly
Anvaya अन्वय · the chain from cause to consequence
Everything above is one platform: a data layer and a loop of agents, each named for what it does. Start at the door that solves your loudest problem; the rest is already installed.
10 · Built by former buyers
We sat in your chair. We know which questions get asked in the boardroom.
Daksh Sahni
Founder. A decade leading people analytics and HR technology, including at Bharti Airtel, Sun Life and Concentrix.
Bring your Monday question. We’ll work it in 30 minutes.
No deck. You bring one real question; we show how the function would answer it, what data it needs, and what the first 90 days look like.