Boston, MA · forward-deployed · product · data · applied AI
Vishnu Kchitti
I turn ambiguous, high-stakes workflows into software people can trust.
At a regional bank, that means replacing fragile Excel, Access, and Alteryx workflows with a governed data platform used across 14 departments. At Johnson & Johnson, it meant working with consultants, business partners, and subject-matter experts to turn competitive-intelligence questions into a production RAG product.
That is also how I think about forward-deployed engineering, legal tech, applied AI, and business automation. Learn the process from the people doing the work. Turn unclear rules into a usable system. Automate the routine parts, while keeping evidence, permissions, and approvals visible.
My open-source work follows the same pattern. I have decoded undocumented wearable data, hardened SQLite replication for a real mobile app, and designed agent tools where models can propose changes but people keep authority.
voice systemThis site was written with a voice profile distilled from seven hours of my own Rhapsode dictation.
Selected work
Different users, same pattern: learn the workflow, find the real constraint, and ship the smallest durable fix.
- Regional bank data platform · zero to one across 14 departments
Selected Microsoft Fabric/Azure and led a three-person migration from Excel, Access, and Alteryx workflows to a governed 343M-row platform. The system reconciles 89M+ transactions penny-perfect and supports 100+ reports.
- add-mcp v2.0.0 · fixed the layer that owned the failure
One setup command broke every Codex chat. I traced it across Firecrawl, Codex, and add-mcp, reproduced it deterministically, and shipped the atomic fix protecting every downstream client.
hybrid config → atomic replacementmerged upstream - liters-mobile · made SQLite replication viable inside a phone
A 153 MB sync could not finish in an iOS background window. I stopped building my own replica when I found a stronger open-source base, then fixed the mobile integration and correctness failures that real use exposed.
two SQLites → onethe lock survives - rhapsode · a signed Mac product I use every day
Dictation designed around one unforgiving constraint: speech should become clean text before I lose my train of thought. It races cloud and local transcription and checks audio before deleting suspected hallucinations.
speech → raced transcription0.6–0.9 s cloud path - noop + noop-cloud · joined an existing product and shipped to its users
I reverse-engineered an undocumented wearable, worked across Swift and Kotlin, and shipped 27 changes upstream. Then I built a permission layer that lets AI agents use the data without silently rewriting it.
raw signal → corroborated data3 bpm vs 42 bpm null - arbiter · an evidence trail for AI-generated claims
A checker accepted 82.8% of generated proofs. After an audit for vacuous and trivial encodings, only 16.3% survived. I built the verification and review layer around that gap.
checker pass → audited proof82.8% → 16.3% - Agent permission systems · useful automation without invented authority
Two products explore the same boundary: what may an agent read, what must a person approve, and what must remain reversible? Health-data edits use propose, confirm, and undo; screen-time access stays enforced by the phone.
read → propose → confirmno silent writes
Experience
I work from user workflow to production system: requirements, architecture, rollout, and the failures that only appear after launch.
- Bristol County Savings Bank — Data Engineer · Taunton, MA · May 2025 – Present
Built and now run the bank's 343M-row enterprise data platform for 14 departments. Selected Microsoft Fabric/Azure, led a three-person migration, and translated reporting and credit rules into governed pipelines, access controls, and seven FDIC classifications across a $2.64B loan portfolio.
- Sorce Jobs — Software Engineer (Contract) · Remote / Boston, MA · February 2025 – June 2025
Worked directly with founders on product strategy, customer personas, IP planning, backend pipelines, and analytics for an early-stage job-matching product.
- Johnson & Johnson Innovative Medicine — Software Engineer · Titusville, NJ / Remote · June 2023 – April 2024
Built a production RAG system for pharma competitive intelligence. Led three engineers and worked with consultants, business partners, and subject-matter experts to turn strategy questions into product requirements for Global Corporate Strategy.
- CRAFT Center, Rensselaer Polytechnic Institute — Research Assistant · Troy, NY · August 2022 – May 2023
Built the data infrastructure behind a published crypto-social-media study, ACM WebSci 2024.
- Avidia Bank — Digital Solutions Intern · Hudson, MA · June 2022 – September 2022
Independently deployed a bank-wide IT asset inventory system for SOX and Massachusetts Division of Banks requirements, improving inventory accuracy by 25%.
The résumé has the full history and bullet-level detail.
Get in touch.
Boston, MA
- Email · vishnukchittibhooma@gmail.com
- GitHub · github.com/vishk23
- LinkedIn · linkedin.com/in/vishnukchitti
If you are building a legal-tech product or using applied AI to automate a messy business workflow, I would like to hear from you. I am especially useful in forward-deployed work: learning the process with the people who run it, automating the routine parts, and turning unclear rules into software they can inspect and trust.