I build backend services and AI agents, and the checks that prove they worked.
Python, Go and TypeScript. Lead engineer at Auteur, a consumer fashion curation startup, where I am the sole engineer. Previously a software engineering intern on Rivian's Data Applications and ML team, building telemetry services and failure monitoring on AWS. M.S. in Computer Science from the University at Albany. Based in New York, NY.
Sole engineer on a platform that curates a catalog spanning many brands around each shopper's own style profile. Built it end to end in Next.js 16, React 19, TypeScript and Postgres: nine screens, an installable PWA, passwordless magic link auth and a nine table Drizzle schema.
Designed a deterministic styling engine that turns a stylist authored rules matrix into per garment outcomes, verified by 446 unit and property tests, with perceptual color matching in OKLab and OKLCH instead of hex comparison.
Built a ranking engine with seven weighted signals, a plain language reason on every result and a diversity pass: 15 ms median to filter 40,000 products, 61 ms to rank and diversify 6,000.
Shipped a Claude vision tagging pipeline constrained to a controlled vocabulary, with confidence gating, a holdout evaluation harness and a review queue that turns expert corrections into training data.
Why the rules stay on the server. The styling rules are the company's core IP, so a build gate scans every shipped JavaScript chunk and fails the build if any of them appear. Catalog ingestion across four affiliate networks runs as a scheduled harvest, so no page load touches an external service or a credential.
Recommendation path
Affiliate catalogsfour networks, harvested on a schedule
Claude vision taggingcontrolled vocabulary, confidence gated
Deterministic styling rulesserver only, checked at build time
Seven signal rankingwith a reason on every result
Diversity passa varied page, not ten of the same
May to Aug 2025 Palo Alto, CA
Software Engineering Intern
Rivian Automotive, Data Applications and ML
Added update and delete to VMCAP, a Go microservice for vehicle telemetry metadata that only supported create and list. The update API is a PATCH with partial updates and protobuf validation, and both ship with CLI commands, so ML teams stopped editing MongoDB records by hand.
Built an event driven failure monitor for AWS Batch simulation jobs, provisioned in Terraform. Failures that used to be silent across 25 to 30 scenario runs now update Simdash and post to Slack.
Fixed 5 production defects by tracing failures from the API layer through PostgreSQL and MongoDB, shipping each fix with a regression test alongside the owning engineers.
Raised test coverage on Polaris, an internal ML dashboard, from 35% to 85% in two sprints with 25+ Cypress tests. Automated Swagger docs in CI and provisioned 3 internal services with Terraform.
Why soft delete. I argued for soft delete over hard delete so a mistaken call could be recovered, and it was adopted in dev and prod. About 80 go tests cover the service's four methods.
Failure path of the Batch monitor
AWS Batchan array job child fails
EventBridge rulematches the failed state change
SQS with dead letter queueno failure event is lost if processing breaks
Lambdamaps the child index back to its scenario
Notify
Simdash status
Slack message
Jan to May 2026 Albany, NY
Graduate Teaching Assistant
University at Albany
Taught CSI 521 Discrete Mathematics and CSI 520 Cryptography to 100+ graduate students, designing assessments and running weekly office hours.
Summer 2023 India
Research Intern, then Software Development Intern
GITAM Deemed University; Frux Software Solutions
Designed an experimental protocol for N dimensional smart contracts and contributed statistical analysis to the research. At Frux, built a Node.js and React pharmacy tracking platform on MongoDB, containerized and deployed to AWS EC2.
Projects
Each one is drawn as the path its data takes. The highlighted step is the decision the project depends on.
Sept 2026
Voice Agent Eval
An evaluation platform for a healthcare voice agent, built in a strict 8 hour timebox. A transcript can sound perfect while nothing changed, so ground truth is the simulated world state, and the headline metric is false completion: the agent said it was done and the world says otherwise.
Comparing two agent versions, headline metrics moved 1 to 3 points while recall on calls that must escalate fell 87.
Why sliced reporting. An aggregate hid the safety regression completely, so regressions are listed before averages by design. Evaluator agreement with human labels went from a Cohen's kappa of −0.05 to 0.86, and a deterministic quote check caught the LLM judge inventing a citation it labeled verbatim.
17 labeled scenariosrefills, scheduling, must escalate cases
Agent v1 and v2run against tools with injected faults
World state as ground truthwhat actually changed, not what was said
Rule evaluators and LLM judgejudge quotes checked against the transcript
Sliced comparisonregressions first, averages second
Oct 2026
Patient Voice Tester
A voice bot that phones an AI medical receptionist and plays a realistic patient, recording both sides of every call and measuring per turn latency, overlaps and dead air.
26 real calls across 23 scenarios produced 143 candidate findings, reduced to 25 issues, 7 high severity, including accepting a mismatched date of birth.
Why the quote check. A finding only counts if its evidence matches the transcript word for word, so claims the model could not ground were dropped.
Both sides recordedlatency, overlaps and dead air per turn
Findings checked against transcriptungrounded claims dropped
Sept 2026
ActionGuard
A deterministic guardrail for a computer use agent operating a claims system. Fourteen rules sort every action into allow, escalate, abstain or block, stopping payments to excluded providers, leaks of patient data and duplicate replays.
Against 26 indirect prompt injection attacks the detector caught 76.9%, yet 100% were contained by verb level authority controls.
Why authority over detection. Detection missed every adaptive semantic attack, so safety rests on what each verb is allowed to do, not on spotting the attack. Standard library only, 83 tests, tamper evident audit trace.
Agent proposes an actiona click, a payment, a data export
14 deterministic rulesno model in the loop
Allow, escalate, abstain or blockauthority set by verb, not by content
Escalation to a named ownera person decides the hard cases
Tamper evident audit traceevery decision recorded
Aug 2026
Fit Score
A physician job match card that shows its own uncertainty. When a criterion is unknown, guessing it misleads and scoring it as zero punishes small and rural sites for missing paperwork, so unknowns are excluded and shown as coverage and a floor to ceiling band.
A site scoring 94 with 43% coverage can be a worse job than one scoring 67 with full coverage, and the band shows it before you read a number.
Process. Synthetic data only, with 43 written design decisions in the repo.
Split known from unknowneach fact labeled by its evidence
Score over known onlyunknowns excluded, never guessed
Coverage and bandfloor and ceiling if unknowns go either way
Match cardscore, reasons, open questions
June 2026 NYC Tech Week
AuditGuard AI
Data integrity triage for manufacturing records ahead of an inspection. Upload a production CSV and a four stage pipeline flags problems with explainable rules, ranks them by audit risk, proposes corrections without touching the original, and produces a report a compliance officer can sign.
Flagged 850+ integrity issues in a 5,000 row manufacturing dataset, each corrected automatically or escalated to a person.
Why the LLM only explains. In a regulated setting a flag has to give the same answer twice, so detection, ranking and every count are plain code. The LLM writes the summary, and it is rejected unless every finding and number in it matches.
FastAPI, Claude API or Groq, React, server sent events. Started as a solo hackathon build.
Summary checked against findingstemplate used if any number fails
Personal project
OpsPilot
An incident triage agent that turns a plain English report such as "login returns 500 after deploy" into a structured change plan, backed by runbook evidence instead of guesswork.
Why a validated plan. Free text is hard to act on under pressure, so the agent must return a change plan as JSON that passes validation, alongside a readable response.
FastAPI, Astra DB vector search, Langflow, MCP, Docker Compose, Nginx
Validated change planstructured JSON, checked before return
Ticket and session memorypersisted for the incident's lifetime
2026 Boston Tech Week
VisionOps
A three agent pipeline that turns freight claim documents into structured facts, built on NVIDIA Nemotron at the Phinite and GMI Cloud event.
Tools
PythonGoTypeScriptFastAPIAWS
Languages
Python, Go, TypeScript, JavaScript, SQL
LLMs and agents
Claude, OpenAI and Groq APIs, Claude vision, tool calling, MCP, RAG and embeddings, structured outputs, evaluation and LLM as a judge, guardrails, LangChain, Langflow