Available for full-time roles · Dallas, TX

Hi, I'm

Prathibha
Devkar.

I'ma Software Engineer

4+ years shipping production ML systems & full-stack applications. MSCS from UT Arlington. I like software that tells the truth about itself, from calibrated confidence scores to honest eval harnesses.

See my work

Resume

GitHub

tip: move your cursor over the network, click to send a pulse, or press⌘K

01About

Building things that just work.

“I find the most satisfaction in building something that just works, reliably and without drama.”

I'm the kind of person who gets genuinely excited about a tricky problem. When something isn't working, I don't just want it fixed. I want to understand why. That curiosity is what pulled me into software in the first place, and four years in, I still learn something new on almost every project.

What keeps me going is the person on the other end. One of the things I'm proudest of is early research I did to help people who can't speak communicate through hand gestures. It taught me that the best work quietly makes someone's day a little easier, which is why I care so much about getting the small details right.

I'm based in Dallas, recently finished my Master's at UT Arlington, and I'm still at UTA as a Research Assistant. I'm happiest on teams that are honest about what's working and what isn't, where people help each other figure things out and nobody has to pretend to have all the answers.

Currently

Research Assistant

SEAR Lab, UT Arlington

Real-time energy monitoring: live HVAC, water-heater and solar telemetry in one dashboard.

Based in

Dallas, TX

--:--:--

local time · CT

years shipping production software

45+

production bugs resolved at Board

96%

gesture-recognition accuracy (Springer, 2021)

02Selected work

Projects with honest numbers.

01Real OTel Tracing

GlassBox

Built a real OpenTelemetry-based tracing SDK for multi-step AI agents — genuine spans on GenAI semantic conventions, a SQLite-backed store, and a waterfall viewer (React + FastAPI). Pointed it, unmodified, at a second project's real agent (RDT Transformer) to root-cause a tool-chaining failure that project had disclosed but never explained: on the second structured-generation attempt, the model doesn't malform its JSON, it collapses into answer-shaped text with no braces at all. Also built a failure classifier reusing confidence-referee's grounding signal unmodified — a third domain — honestly reporting its current 33% recall rather than inflating it.

PythonOpenTelemetryReactFastAPI

02~47% Fewer Params

Recurrent Depth Transformer (RDT)

Implemented a Recurrent Depth Transformer from scratch in PyTorch, replacing stacked transformer layers with a single recurrently-looped block. Matched a 6-layer GPT baseline's multi-hop accuracy (1-5 hops) with ~47% fewer parameters; running extra loops at inference (6/10/14) did not extend accuracy past the trained depths. Built a tool-calling agent on top of it, caught a misleading evaluation metric reporting 100% completion regardless of correctness, and rewrote the harness to report results honestly.

PythonPyTorchTransformersEvaluation

0358% vs 35% Precision

Confidence-Aware Meeting Intelligence

Built a meeting summarizer and action-item extractor that scores the reliability of each output instead of presenting everything with equal certainty. A local LLM (Ollama, llama3.2:3b) extracts summaries and action items, independently cross-checked by a fine-tuned DistilBERT classifier; a calibrated confidence score (logistic regression on quote-grounding, owner-attribution plausibility, cross-model agreement, and self-reported confidence) was trained on 228 hand-labeled examples. The confidence layer has since been extracted into a standalone, reusable package — see confidence-referee below.

PythonOllamaDistilBERTStreamlit

040.97 AUC, 2nd Domain

confidence-referee

Extracted the confidence-calibration layer out of the meeting-intelligence project into a standalone, reusable Python package: domain-agnostic signals (grounding, cross-model agreement, self-reported confidence, an optional domain validator) calibrated per-domain via logistic regression. Validated the abstraction actually generalizes — not just in theory — on a second, unrelated domain: RAG-answer hallucination detection on HaluEval, using one signal alone, completely unmodified.

Pythonscikit-learnLibrary DesignCI/CD

050.97 AUC (Win)

Apex F1 Predictor

Full-stack ML app forecasting Formula 1 race outcomes from historical race data (2023+) with a RandomForestClassifier and a FastAPI backend, including autonomous retraining after each race and WebSocket-based live telemetry. Reached AUC of 0.97/0.94/0.84 for win/podium/points-finish probability on a chronological hold-out of the most recent races; exact finishing-position prediction is harder (~3.6-position MAE), a limitation surfaced directly in the app rather than hidden.

PythonFastAPIscikit-learnWebSocket

03Toolbox

The stack I reach for.

PythonTypeScriptJavaScriptC#SQLPyTorchTensorFlowKerasscikit-learnOpenCVNumPyPandasOllamaLLM EvaluationConfidence CalibrationDistilBERT Fine-TuningPrompt EngineeringMCPASP.NET

Node.jsFastAPIRESTful APIsReactReact NativeAureliaD3.jsHTML5CSS3PostgreSQLMySQLAWSGCPDockerKubernetesOpenTelemetryGit/GitHubAzure DevOpsJira

hover a cluster · drag any node

Languages

PythonTypeScriptJavaScriptC#SQL

04Research

Published work.

Springer · 2021

Speaking Mouth System for Speech-Impaired People Using Hand Gestures

Advances in Intelligent Systems and Computing, vol 1163. Springer, Singapore

Developed a webcam-based hand gesture recognition system using a CNN-based pipeline.
Achieved 96% recognition accuracy using ROI-based preprocessing and HSV image processing.
Received KSCSTE INNOVATE research funding for project excellence.

Read paper

05Experience

Where I've shipped.

Apr 2026 - Present
Research Assistant
SEAR Lab, UT Arlington · Arlington, TX
Current
- Refactored a 6-service Docker stack (Mosquitto MQTT, InfluxDB, Grafana, a FastAPI Data Hub, Python collectors) into a single self-contained FastAPI application with an embedded dashboard, eliminating the Docker/Grafana/MQTT infrastructure dependency entirely.
- Integrated live telemetry from three hardware APIs — LG ThinQ (AC), Rheem EcoNet (water heater), and Enphase (solar) — into a unified REST API refreshed every 60 seconds by a background asyncio collector, behind a dashboard that charts real-time net load (solar minus HVAC/HWH draw).
May 2025 - Mar 2026
Software Developer Intern
Board Americas · Remote, Texas
- Resolved 45+ medium-to-high severity production bugs over 11 months across an enterprise forecasting platform (Aurelia/TypeScript, C#), improving platform stability across UI, state management, and backend API layers.
- Diagnosed and fixed critical D3.js charting bugs, including fiscal-year x-axis display errors and incorrect actual-vs-forecast rendering, correcting chart data clients relied on for business decisions.
- Fixed a company-wide cache-invalidation bug that let forecast data drift out of sync across users, by changing invalidation scope from per-user to per-company; also resolved a null-date bug across all datepicker components that was triggering 500 errors in production.
Oct 2019 - Nov 2022
Technology Analyst
Infosys Private Ltd. · Trivandrum, India
- Developed React Native UI and integrated backend APIs for a curbside-pickup feature added to the Tractor Supply Company mobile app, improving user satisfaction by 30% based on pilot-release feedback.
- Built RESTful APIs in Node.js for an e-learning platform supporting 200+ concurrent users across student, teacher, and admin roles, enabling real-time session scheduling, participation, and new-user registration.
- Designed PostgreSQL schemas and optimized SQL queries for the platform; built a JMeter load-testing pipeline simulating 1K-10K virtual users, which identified and resolved a calendar API bottleneck for a team-reported 40% improvement in data-retrieval speed.
- Deployed the platform to live servers via PuTTY/WinSCP and managed process reliability with PM2, ensuring stable delivery across environments.
3× Infosys Insta Awards (2021–22): JMeter performance pipeline, Meridian events platform, organizing a team event

Certification

IBM Data Science Specialization

IBM · Coursera · May 2024 · 5 courses

Verify