Forward Deployed AI Engineer · Bentonville, Arkansas · open to remote US
Embedded with your team. Shipping the system myself. Measured in revenue, not hours.
Four industries, one method: discovery with the people doing the job, a spec they can read, the build, and a number afterwards. The number I care about is revenue. Deals closed 25% more often when my platform was part of the pitch 11, $2M+ in client payments arrived sooner 10, and the clients on my integrations kept $2.3M a year 3. If I cannot measure it, I say so.
Most recently I built a dental-clinic SaaS alone with Claude Code 1,4. Two clinics run on it. There is no third until the legal work is finished, because that is the right order.
Indian dental clinics run on a paper diary, a tooth chart and a bill book. Dantra replaces all three under the clinic's own name.
Independent dentists in India. My cofounder, Dr. Reshma Abraham, is one of them and runs the clinical side.
Wrote the specs and eight architecture decisions, then built and shipped the production system myself with Claude Code: records, tooth charting, treatment plans, GST billing, scheduling, WhatsApp messaging 1,4.
One clinic can never see another's data, and the database enforces that, not the app 5. Bills are gap-less and cannot be edited after the fact 6. Every change to a clinical record is logged.
An adversarial AI security review of the codebase: zero critical findings 7. A cross-tenant isolation test matrix that blocks every release 5.
Enterprise shippers wanted freight data inside their own systems. I owned each integration from the first workshop to go-live.
Accounts like Verizon, Ecolab and ESCO, and internally 300+ branches reading different numbers from different reports 2.
Ran discovery with customer teams, wrote the EDI and API specs, tested the integrations myself, and took more than 30 live on schedule 8. Built the Power BI and Azure Synapse platform that became the one source of reporting 9.
98 of 100 transactions on automated lanes flowed with no manual touch, about $2.3M a year saved for those clients 3. Ad-hoc report requests fell 40% 9. $2M+ in client payments came in sooner 10. Deals with the platform attached closed 25% more often 11.
“What stood out was how she worked with customers. She ran discovery directly with their teams, wrote the specifications, tested the integrations herself, and stayed with the account until it was working. She is technically strong, commercially aware, and calm under pressure, and she raised the standard of the analysts around her.”
A finance team typing orders and journal entries in by hand got about 160 hours a month back.
Software robots on order entry and journal postings: process time down 98%, accuracy up 16% 13. Then 85% of reporting moved from Excel to automated dashboards, over 15 months 12.
“What I valued most was her thoroughness in every aspect of her work. She was diligent about ensuring data made sense to stakeholders and internal customers, rigorous in validating results for accuracy, and thoughtful in designing automation that delivered value seamlessly across the organization.”
Reads an invoice, cites the page for every field, and waits for a named person to approve in Slack. 10 of 10 test invoices correct at $0.022 each. 16
Checks an AI call against the per-call cap, the run budget and the client's monthly budget before making it. Settles the real cost after. 17
The pattern from Dantra as a small TypeScript library: drafts only from today's data, strict JSON, spend cap checked first, nothing saved until a person approves.
A production-shaped warehouse on DuckDB with 86 tests. Its 26 business questions double as the eval set for an analytics agent. 18
A person approves anything that sends, saves or pays.
Claude drafts. It does not decide. In Dantra and in every pipeline I have published, the approve step is the only path that moves anything.
Every AI step ships with an eval set and a cost per run.
Ten hand-checked cases beat a demo. If I cannot say what a run costs, it is not ready.
"Saves time" should be a number.
Hours a month, dollars a year, requests that stopped coming. If I cannot measure it, I say so.
Do it in the right order.
Two clinics are on Dantra. The third waits for the legal work, not the other way round.
If your team still does something by hand every day, tell me about it.
- 1Dantra, dantra.in. Staging and production on Cloudflare Workers. 13 prototype iterations, 8 architecture decision records, 2026. Two clinics in production. ↗
- 2Mainchain, Mainfreight. Adoption passed 80% of 300+ branches across NZ, Australia, Europe and Asia, 2021 to 2023.
- 3EDI and API integrations at Mainfreight: 98% straight-through processing on automated lanes, about $2.3M a year saved for enterprise clients including Verizon and Ecolab.
- 4Built solo with Claude Code multi-agent workflows. TypeScript, Next.js, Cloudflare Workers, Postgres.
- 5Postgres row-level security on every tenant table. CI is blocked by a cross-tenant isolation test matrix with no override.
- 6Gap-less sequential invoice numbers per branch, credit-note voids, money math on the server only, concurrency-tested against real Postgres.
- 7Adversarial AI red-team review of the production codebase, 2026: zero critical findings.
- 8More than 30 enterprise EDI and API integrations taken live on schedule, as stated by Nathan Thomas, who managed the work. Partners included project44, FourKites, E2open, Flexport and Infor Nexus.
- 9Power BI platform and semantic layer over an Azure Synapse warehouse, 15+ dashboards, ad-hoc report requests down 40%.
- 10Accounts-receivable aging dashboard: $2M+ in client payments pulled forward.
- 11Team of 4 analysts and developers. Deals with the platform attached closed 25% more often; team productivity up 25%.
- 12Modern Gourmet Foods: 85% of reporting moved from manual Excel to automated Power BI and SQL, manual effort down 40%.
- 13Kofax RPA on order entry and journal entries: process time down 98%, accuracy up 16%, about 160 hours a month returned. Nadeem Mumal cites the same 160 hours.
- 14TCS for Citi, 2012 to 2016: requirements, specs and the data platform behind a global payments business, 100+ stakeholders in 5 regions.
- 15Scrum lead across 5 regions; feature delivery time down 25%.
- 16claude-invoice-intake. Live eval on 2026-10-13: 10 of 10 test invoices with every field and citation correct, $0.022 per invoice. 63 tests run without an API key. ↗
- 17cost-guard. Per-call cap, run budget and monthly client budget, with a ledger that reserves before the call and settles after. 15 tests, standard library only. ↗
- 18dbt-logistics-warehouse. Star-schema marts on DuckDB, incremental fact, SCD2 snapshot, 86 data tests, CI. The 26-question bank doubles as the eval set for an analytics agent. ↗
- 19LinkedIn recommendations from Nathan Thomas (September 11, 2026) and Nadeem Mumal (September 9, 2026), both of whom managed Aishwarya directly. Quoted verbatim. ↗