Himanshu Srivastava
Data and analytics engineering, mortgage and financial services
Dallas, TX · open to data analyst, analytics engineer and data engineer roles
Most of my work is automation in places where being wrong is expensive: mortgage GL reconciliation, Ginnie Mae pool delivery, a CI/CD platform at a regulated lender. The through-line is a preference for systems that refuse to guess, and that hand the decision back to a person at exactly the point where a person is needed.
5releases on PyPIcustody-ledger
39,130records readthree national registers
457tests passingpytest --collect-only
0applications sent by the machinestructural, not a setting
Dallas, TX · two master’s degrees · four years across analytics and engineering
I am a data and platform engineer working in mortgage technology. At Supreme Lending the systems I build touch loan files that regulators read, and that context is most of why I work the way I do: where being wrong is expensive, a system that says “I cannot prove this” is worth more than one that produces a confident number.
Before the move into engineering I spent four years in analytics delivery and business analysis — requirements, stakeholders, reporting, and the unglamorous work of making numbers agree across systems that disagreed by design. The two halves are the point: I can sit with the person whose process it is, and then go build the thing.
Outside work I build the same way. Everything in the next section is mine, including the parts that did not work — which are published too.
Every figure below names the command that produced it.
I ship code people actually install.
A mortgage lender has to show a regulator what its model produced and what a human did about it. A log asserts that. A hash chain with signatures proves it, and an anchor scheme makes deletion from the end detectable.
5releases on PyPIpypi.org/project/custody-ledger
151tests, 15 filespytest --collect-only
2026-09-05
1runtime dependencypyproject.toml
2backends, both in CISQLite and Postgres
I say nothing the data cannot support.
Reads three national medicines registers and turns them into named, dated triggers. Tanzania publishes neither expiry nor reliable dates, so it produces nothing. That zero is the rule working, not the pipeline failing.
39,130records readNAFDAC, ANVISA, TMDA
hashed snapshots
6,899triggers producedeach carries a date
0from Tanzaniano dates published,
so nothing dated claimed
10test files, two gatesls tests/
I automate work that costs real hours.
Every sentence on a generated resume traces to a recorded claim. A containment gate discards generated prose whole if it introduces a figure, a proper noun or a seniority the evidence does not carry. It once scored a posting 98/100 with “no gaps”, because the requirements it did not understand were invisible rather than unmet. That bug is why the gate exists.
457backend tests, 60 filespytest --collect-only
2026-09-03
88API endpointsgrep -c '@app.' app/main.py
281commits, two reposgit rev-list --count
0applications sent by the machinestructural, not a setting
I publish the results that did not work.
Scores short-dated options strategies net of costs, with bootstrap confidence intervals, and calls a strategy unproven when the interval spans zero. Seven were measured. None cleared. The table is published as it came out.
801tests, 82 filespytest --collect-only
2026-09-03
7strategies measuredbootstrap CIs, net of costs
0provenevery interval spans zero
215commitscode private, demo real
Supreme Lending, production systems that touch loan files regulators read.
A process that lived in one person’s head now lives in sixteen rules.
A specialist spent two to three hours a day cross-referencing Power BI, the loan origination system and the A&B ledger to classify and route mortgage GL discrepancies. Nothing was written down, so the first work was structured interviews and sitting through live sessions. The second was writing it as rules that could be tested.
2–3 hrsdaily, beforeone specialist, measured
100+open cases routedby debit/credit class
16rules codifiedfrom interviews
3systems joinedPower BI, LOS, A&B
Three loan systems, one valid submission.
ULDD XML exports from three different origination systems, merged into valid Ginnie Mae PDD XML for SFPDM submission, with schema errors corrected automatically rather than bounced back to a person.
$18.2MUPB, first deployment4 pools, 47 loans
3source systemsEncompass, MeridianLink,
Blue Sage
ULDD → PDDschema-validated, errors corrected in placeMISMO schema checked before anything moves downstream
The pipeline cannot grant itself permissions.
CodeQL, secret scanning and dependency review, run in parallel rather than in series. A critical-severity CVE fails the build. The pipeline holds no write access to IAM, because a regulated lender cannot run a deployment system that can widen its own access.
18 → 9minutes, PR checksparallelised the stage
5pipeline stages ownedsecurity stage
3files to onboard a teamreusable workflow contract
0IAM write accesshumans provision identities
Three rules that hold across all of it
Every claim a system of mine makes traces to a record. A number that is not in the record cannot be checked by anyone, so it does not get written.
Not knowing what a term means is never a reason to assume it is satisfied. Scoring that flatters itself is worse than no scoring.
The last step before anything irreversible belongs to a person. That is a structural property in what I build, not a setting somebody can switch off.
Analytics and data engineering