Transitioning to data developer role

Prayas Gautam

MSc Computer Science at Kansas State University, 4.00 GPA. Modernized federal engineering tools, internal banking applications, and AI systems that ship.

Download résumé prayasgautamm@gmail.com LinkedIn
0.00
Graduate GPA
0K+
Lines of VB6 ported
0%+
Numerical parity, legacy engine
1 wk2 hr
Docs pipeline turnaround
01 / selected work

Things I've built, and why they were hard

Most of this is federal or bank-internal, so there's no public repo to link. Open any card for the write-up and a diagram you can drive yourself.

01

Rebuilding a federal dam-safety tool for 64-bit

10K+ lines of legacy VB6 rebuilt in C#/.NET, with the numbers proven to still match.
Graduate Research Assistant · Kansas State University · Dept. of Computer Science Jan 2025 – Present
WinDAM C · embankment overtopping & breach
01 reservoir pool02 crest03 upstream face04 downstream face

WinDAM C and EFH-2 are USDA/NRCS tools that model how an earthen embankment overtops and breaches. Both were Visual Basic 6, 32-bit only — a language out of support for years, on a toolchain with no future on modern Windows.

  • Engineers rely on continuity of results, so a rewrite could not simply produce different numbers and call them better. Parity against the legacy engine had to be demonstrated, not assumed.
  • The existing engineering input-file format had to keep working; users have files going back years.
  • Legacy VB6 I/O was brittle, so the data layer had to be replaced rather than ported line for line.
  • Ported 10K+ lines of VB6 to C#/.NET on WinUI 3 around an MVVM architecture, owning the design through stakeholder review.
  • Designed a new data layer that parses the engineering input files, serialises structured state to JSON, and reloads it into the simulation engine, replacing the brittle legacy I/O.
  • Validated numerical parity against the legacy 32-bit application using numdiff across 10+ regression test cases.
  • Used Claude to generate synthetic input files that pushed regression coverage into edge cases, reviewing every generated case before it entered the suite.
95%+
Numerical parity with the legacy engine
10K+
Lines of VB6 ported
10+
Regression cases validated with numdiff
C#.NETWinUI 3MVVMVB6 migrationnumdiff
02

Killing a week of manual documentation per release

Documentation that regenerates itself on every merge, cutting a week of manual work to under two hours.
Technology Intern, Software Development · INTRUST Bank · internal platform May 2026 – Aug 2026
AI-assisted documentation pipeline

Technical documentation across ten-plus internal applications was written by hand. It drifted from the code almost immediately, and a full turnaround took about a week.

  • It had to run inside the existing CI, not as a separate system somebody would have to remember to trigger.
  • The goal was zero manual authoring. A pipeline that still needed a human to write the prose would have re-created the original problem.
  • Built the generation and sync logic as C# backend services, driven from GitHub Actions CI/CD.
  • Wired it to run on merge, so documentation regenerates as part of shipping rather than after it.
  • Extended coverage across the full set of internal applications rather than a single pilot repo.
1 wk → 2 hr
Documentation turnaround
10+
Applications synced on merge
Zero
Manual authoring steps
C#GitHub ActionsCI/CD
03

An assistant for engineers dropped into unfamiliar code

An assistant that answers from the documentation the pipeline above generates.
Technology Intern, Software Development · INTRUST Bank · internal platform May 2026 – Aug 2026
Internal assistant · context & answer path
Solid path shipped · dashed legs scoped, not yet built

Engineers picking up an unfamiliar internal application had no fast way to get oriented. The knowledge existed, but finding it meant interrupting whoever happened to know.

  • A bank environment, so the assistant had to sit inside the organisation’s existing Azure footprint.
  • Answers had to be trustworthy enough to act on, which made evaluation part of the build rather than an afterthought.
  • Cost mattered as much as capability. An internal tool used casually all day cannot be priced like a frontier demo.
  • Fed it the docs the pipeline above already produces — passed straight into the system prompt as context, not retrieved at query time. The two projects are one system: the first is what makes the second useful.
  • Ran the same workload against GPT-5.6 Terra and GPT-4o mini, and shipped 4o mini: materially cheaper, with a context window big enough to hold the doc set. The larger model was not buying enough to justify the cost.
  • Dependency-injected clients kept the model layer swappable, which is what made that comparison cheap to run in the first place.
  • Owned prompt design and evaluation, treating answer quality as something to measure rather than eyeball.
  • Scoped Azure AI Search-backed RAG and MCP tool-calling as the next integration path: designed for, not yet shipped. Retrieval becomes worth it once the doc set outgrows the context window.
Deployed
And adopted by the team in use
GPT-4o mini
Chosen over GPT-5.6 Terra on cost
RAG + MCP
Scoped as the next integration
Azure AI FoundryGPT-4o miniRAGMCPC#
04

Proving the retriever, not the model, was the bottleneck

A retrieval pipeline whose real finding was that retrieval, not the model, was the ceiling.
Course project, three-person team · CIS 732 Machine Learning · Kansas State University Spring 2026
SciFact · retrieval and verification
Macro F1, 5-fold cross-validation, SciBERT

SciFact asks whether a scientific claim is supported, refuted, or unsettled by the literature. That needs two things: finding the right evidence among 5,183 abstracts, and judging whether it backs the claim. Most reported results measure only the second.

  • 809 training claims is small for fine-tuning a transformer, so any single split proves nothing. Every comparison had to run through 5-fold cross-validation on a fixed seed before it meant anything.
  • The three labels were badly imbalanced, so class weights had to be recomputed per fold rather than fitted once globally.
  • DeBERTa-v3 returned NaN losses in its first epoch under fp16 and only trained once forced to fp32. A silent numerical failure with no error to read is far slower to find than a crash.
  • A BM25 index over all 5,183 abstracts retrieves candidate evidence sentences, which feed a transformer NLI verifier fine-tuned on claim and evidence pairs.
  • SciBERT and DeBERTa-v3-base were fine-tuned with the full encoder unfrozen, then each evaluated twice: on gold evidence, and on what BM25 actually retrieved. Running both is what separates retrieval quality from verifier quality — and why the project ended with a finding, not just a score.
  • Everything was compared against a matched TF-IDF and logistic regression baseline using paired t-tests across the same fixed folds.
  • Post-hoc temperature scaling was fitted on development-set logits by L-BFGS, giving T = 1.2507. Mild overconfidence, and monotonic, so calibrating the probabilities never changes which label is predicted.
0.694
Macro F1 on gold evidence, p = 0.005 vs baseline
0.531
Same verifier on BM25-retrieved evidence
16–34
Macro F1 points lost to retrieval
PythonHuggingFaceBM25SciBERTDeBERTaNLICalibration Source on GitHub
Also built

Academic and personal work — no confidentiality, so described in full.

EV Charger Finder

React · TypeScript · Node.js / Express · ArcGIS API

A full-stack geospatial web app with a RESTful Express API and the ArcGIS API behind location-based search, integrating third-party APIs to answer queries in real time.

QuickCommerce

PostgreSQL · Supabase · React · Vite

A normalised PostgreSQL schema on a Supabase backend with a React frontend. Relational entities were modelled from scratch and every CRUD workflow validated end to end.

02 / experience

Where I've worked

Technology Intern, Software Development

May 2026 – Aug 2026
INTRUST Bank · Wichita, KS
  • Developed internal banking applications in C#/Blazor on .NET using a vertical slice architecture, delivering production features in two-week Agile sprints with peer code review.
  • Built an AI chat assistant on GPT-4o mini in Azure AI Foundry, deployed and adopted by the team for questions about unfamiliar internal apps. Owned prompt design and evaluation; scoped RAG and MCP tool-calling next.
  • Automated an AI-assisted documentation pipeline in C# on GitHub Actions CI/CD: docs regenerate and sync across 10+ internal applications on every merge, no manual authoring. Turnaround fell from a week to under 2 hours.
  • Authored two internal data-governance guides for a federated platform: CDC ingestion, medallion layers on ADLS/Apache Iceberg, cataloging, lineage, PII classification and data-quality SLAs.
Case studies Killing a week of manual documentation per release An assistant for engineers dropped into unfamiliar code
C#BlazorAzure AI FoundryGitHub ActionsRAG

Graduate Research Assistant

Jan 2025 – Present
Kansas State University, Dept. of Computer Science · Manhattan, KS
  • Leading 64-bit development of WinDAM C and EFH-2, legacy USDA/NRCS engineering tools, porting 10K+ lines of legacy VB6 to C#/.NET/WinUI 3 on an MVVM architecture, owning design through stakeholder review.
  • Validated numerical parity against the legacy 32-bit application with numdiff across 10+ regression cases, matching to 95%+ — with better floating-point precision in several scenarios.
  • Used Claude to generate synthetic input files expanding regression coverage to edge cases, reviewing every generated case before adding it to the suite.
  • Designed the application data layer that parses engineering input files, serializes structured state to JSON, and reloads it into the simulation engine, replacing brittle legacy VB6 I/O.
Case studies Rebuilding a federal dam-safety tool for 64-bit
C#.NETWinUI 3MVVMVB6 migrationnumdiff

Graduate Assistant

Jul 2025 – Aug 2025
Kansas State University, Dept. of Psychological Sciences · Manhattan, KS
  • Applied Python, CVAT and AI-assisted workflows to automate annotation validation and data cleaning for behavioral datasets used in model training, replacing a fully manual review process.
PythonCVAT

Business Development & IT Officer

Jul 2024 – Dec 2024
Jaya Consulting Services · Kathmandu, Nepal
  • Worked directly with clients to scope requirements, trained staff on applying AI tools to research tasks, and organized team data workflows, improving turnaround across consulting projects.
Agile
03 / skills

What I work with

Languages
JavaScriptSQL
Cloud & AI
LLM integrationPrompt engineering & evaluation
Frameworks & tools
Entity Framework CoreRESTful servicesGitDockerVisual StudioCursorGitHub Copilot
Databases
MySQL
Concepts
Data structures & algorithmsOOPConcurrency & parallel processingVersion control & technical documentationDebugging & testing
Tap an outlined skill to filter the experience above
04 / education

Education & service

Jan 2025 – Dec 2026

MSc Computer Science

Kansas State University · Manhattan, KS
GPA 4.00 / 4.00
Jul 2019 – Jul 2023

BE Computer Engineering

Goa College of Engineering · Goa, India
GPA 3.62 / 4.00
Leadership & service

Graduate Student Leader, AI Task Force

Kansas State University

Steering-committee member shaping university-wide guidance on responsible Generative AI use for graduate research.

05 / contact · graduating december 2026

Data, software, AI, and the business around them. Let's talk.

LinkedIn Résumé (PDF)
Prayas Gautam Manhattan, KS