Rudraksh Syal

Data Science Engineer

SeniorAI/ML Engineer

11years in tech

5years in AI/ML

7/10State Management & Data Flow

Overview

Senior frontend UI engineer specializing in device-first web apps and voice-enabled remotes for TVs and mobile PWAs. Their strongest proven skill is resilient async state and voice interaction engineering, demonstrated by the phone hold-to-talk state machine and the Playwright test suite that reproduces and validates iOS/Safari STT edge cases (grokbot_tv_launcher/relay/phone.html and tools/test-phone.js). What is not evidenced is a large-scale component library or multi-team design-system ownership and limited direct React/Next production code authored by the developer in the human-authored files.

Timeline

Data Science Lead · Lead

Hana E Solutions · Full-Time

Oct 2026 to Present 1 Month San Jose Partially remote

Built agentic AI decision workflows from scratch for client engagements, starting with pricing use cases. Worked on early proof-of-concept design and evaluation to translate decision requirements into deployable logic. Partnered with stakeholders to shape the initial scope and validate feasibility.

AI Agents

Senior Data Scientist · Senior

Vista · Full-Time

Oct 2021 to Jul 2026 4 Years 9 Months San Jose Partially remote

Led marketplace pricing and decision systems, including replacing static pricing bands with a supply-demand equilibrium engine. Owned forecast-driven MSRP recommendations and improved forecast accuracy using automated daily reporting. Built XGBoost-based lead propensity scoring integrated into Salesforce and developed LLM-derived cancellation-policy badges for operator contracts.

XGBoost

Salesforce

LLM

Senior Data Scientist · Senior

84.51° · Full-Time

Jul 2019 to Oct 2021 2 Years 3 Months Cincinnati In office

Designed Bayesian structural time-series causal measurement using synthetic controls to evaluate retail merchandising impacts. Communicated estimated effects across sales, visits, and household metrics to support go/no-go decisions. Conducted store segmentation experiments to recommend selective changes while monitoring for meaningful deterioration in key outcomes.

Time Series

Purdue University

Master's Degree · Business Analytics and Data Science

2019 West Lafayette, Indiana

Data Scientist · Middle

Practo Technologies · Full-Time

Jun 2015 to Mar 2018 2 Years 9 Months Bengaluru In office

Built reporting applications with Shiny (R) and Flask to replace an external analytics platform. Applied NLP to analyze customer feedback and used A/B testing on product journeys to identify conversion drop-offs. Converted insights into actionable recommendations for improving funnel performance.

Shiny

Flask

A/B Testing

NLP

Technical skills

Languages

Python

JavaScript

Node JS

SQL

Python

Flask

Requests

Pydantic

Asyncio

AI/ML

LLM

XGBoost

NLP

Pandas

NumPy

Scikit-learn

LangChain

OpenAI SDK

Gradio

MLFlow

Frontend

React.js

Next.js

Tailwind CSS

Chart.js

Other

Shiny

A/B Testing

Vite

Dotenv

Amazon SageMaker

ChatGPT

Time Series

ESLint

Snowflake

AI Agents

Spark

GitHub

Tableau

Senior AI/ML Engineer Confidence: Medium LLM Engineer

LLM engineer at a senior level specializing in building agent orchestration and human-in-the-loop automation systems. The strongest proven skill is architecting and implementing agent evaluator loops, middleware and session-managed tool ecosystems as shown by the Sidekick implementation (the_sidekick_agent/sidekick.py and app.py). There is little or no public evidence of custom model training, GPU/quantization optimization, or large-scale production model serving and monitoring pipelines.

Model Architecture & Training

1/10

How well models are designed and trained

Little to no custom model architecture or training; work uses hosted LLMs and SDKs rather than building or tuning models.

Evidence

the_sidekick_agent/sidekick.py: ChatOpenAI evaluator instantiation (.evaluator = ChatOpenAI(...).with_structured_output)

ai-digital-twin/app.py: openai.OpenAI usage for chat completions (openai.chat.completions.create)

evidentML/pyproject.toml: lists mlflow and scikit-learn as dependencies but no custom training loops present

Data Pipeline & Feature Engineering

5/10

How data is prepared for models

Concrete, well-structured tabular data engineering: dataset profiling, deterministic quality checks, safe cleaning and a synthetic benchmark generator with deliberate planted issues.

Evidence

evidentML/src/evidentml/data_engine/profiler.py: profile_dataset and profile_column implementations

evidentML/src/evidentml/data_engine/cleaning.py: clean_mixed_numeric_column, clean_categorical_variants and apply_safe_cleaning

benchmarks/generators/generate_binary_classification_v1.py: realistic dataset generator and inject_data_quality_issues with ground-truth

Experimentation & Evaluation

4/10

How results are measured and tested

Basic experimentation and evaluation scaffolding with a ground-truth benchmark and unit tests; reproducible scoring is present but limited in scale and tooling integration.

Evidence

evidentML/src/evidentml/evidence/benchmark.py: score_audit / expected_findings extraction

evidentML/tests/test_benchmark_evidence.py: unit test for audit scoring

benchmarks/generators/generate_binary_classification_v1.py: write_ground_truth creates the benchmark answer key used by evaluations

MLOps & Deployment

4/10

How models are shipped to production

Practical MLOps and deployment prototypes: Gradio apps, agent lifecycle management, session handling and clean shutdowns; lacks production-grade serving/monitoring pipelines.

Evidence

the_sidekick_agent/app.py: Gradio app wiring and lifecycle callbacks (setup, reset, free_resources)

the_sidekick_agent/sidekick.py: setup and cleanup for MCP sessions, resume/pause logic and evaluator loop

ai-digital-twin/app.py: iterative tool-call handling loop with OpenAI SDK integration

Computational Efficiency

2/10

How efficiently computing resources are used

Some runtime controls and streaming usage (rate/limit guardrails) but no low-level efficiency engineering, GPU optimizations or quantization work.

Evidence

the_sidekick_agent/sidekick.py: ModelCallLimitMiddleware usage and streaming worker.astream handling

the_sidekick_agent/sidekick.py: TolerateToolErrors middleware to prevent tool failures from crashing runs

Research Depth & Innovation

3/10

Depth of research and new ideas

Thoughtful agent orchestration and human-in-the-loop evaluation loop provide solid engineering innovation for agent workflows, but no novel ML algorithms or paper-level reproductions.

Evidence

the_sidekick_agent/sidekick.py: custom evaluator loop that retries worker runs with structured feedback and handles pauses for human approval

the_sidekick_agent/5_lab5.ipynb: design notes and demonstration of middleware stack and evaluator-driven iteration

Expertise

AI Agents & Agentic Workflows· Senior

Technologies

SQL

Snowflake

LangChain

Spark

ChatGPT

MLFlow

XGBoost· since 2021

Scikit-learn

AI Agents· since 2026

NLP· since 2015

Gradio

OpenAI SDK

Pandas

NumPy

LLM· since 2021

Shiny· since 2015

Time Series· since 2019

Amazon SageMaker

GitHub

ChatGPT· mentioned only

Recommendations

Develop end-to-end agentic workflows and human-in-the-loop automation systems, particularly integrations that require robust pause/resume and evaluator feedback.
Build deterministic data-quality and benchmarking tooling for tabular ML use cases, including ground-truth generation and automated audit scoring.
Prototype interactive demos and Gradio-based dashboards for internal tools and small-scale user studies that exercise tool-calls and evaluator logic.
Extend the benchmarking/evidence system toward experiment tracking (MLflow) and clearer CI integration for repeatable evaluation runs.

Repositories

The developer's experience in this domain has been verified based on AI analysis of the following repositories:

evidentML the_sidekick_agent ai-digital-twin

Senior Backend Developer Confidence: High API Engineer

A senior backend engineer focused on building resilient, real-time consumer-device relays and LLM-backed assistant integrations with strong practical judgment. The strongest proven skill is robust LLM orchestration with tool-calling, budgeted fallbacks and resume/retry behavior, evidenced by relay/brain-llm.js together with relay/metrics.js and the test suites (tools/test-phone.js, tools/test-nav.js). There is limited public evidence of multi-node database migrations, large-scale distributed coordination, or a formalized enterprise-grade CI/CD/multi-region operations setup.

API Design

6/10

How well APIs are designed

API design is pragmatic and purpose-built: clear function/tool schemas for LLM function-calling, a small HTTP relay with SSE and well-scoped endpoints; lacks an explicit versioning/long-term idempotency strategy.

Evidence

grokbot_tv_launcher/relay/server.js: routes and SSE endpoints (/events, /ask, /stt, /tv/show)

grokbot_tv_launcher/relay/brain-llm.js: TOOLS array with explicit function schemas (search_titles, get_ratings, play_on_tv, launcher_action)

grokbot_tv_launcher/relay/evals/run.js: in-process eval runner exercising the /ask pipeline and flags (--no-judge, --brain)

Data Layer & Database

3/10

Working with databases

Data layer is lightweight and practical (local JSON stores, local chat persistence, simple budget file); there is no evidence of a relational migration history or multi-node transactional DB design.

Evidence

grokbot_tv_launcher/relay/chats.js: file-based chat persistence and chat primitives (create, save, load)

grokbot_tv_launcher/relay/brain-llm.js: budget persistence to logs/llm-budget.json and seedFromLog() reading metrics.jsonl

ai_portfolio/worker/index.ts: Cloudflare Worker entry showing D1 binding awareness (config for edge DB) but no migration chain

Scalability & Performance

5/10

Handling load and speed

Shows deliberate caching and background-refresh (stale-while-revalidate), prefetching and LLM-budget based rate-limiting; not a full distributed scaling plan but good pragmatic performance attention.

Evidence

grokbot_tv_launcher/js/discover.js: refreshLive(), platStale()/refreshPlatform(), cacheArt() and prefetchPlatform() (stale-while-revalidate and prefetch)

grokbot_tv_launcher/relay/brain-llm.js: budget(), budgetBlock(), endpoints() and fallback logic including local Ollama detection (ollamaUp())

grokbot_tv_launcher/relay/metrics.js: metrics hooks used to time tools and record latencies (M.timeTool, metrics collection)

System Architecture

5/10

Overall system structure

Clean modular separation (relay server, LLM brain, tools, TV client and phone UI, test harness, separate Python agent project). Architecture is pragmatic and well-factored for a single-product stack rather than a multi-service enterprise platform.

Evidence

grokbot_tv_launcher/relay/server.js: orchestrates session, SSE and TV-specific routes separate from LLM logic

grokbot_tv_launcher/relay/brain-llm.js: LLM orchestration and tool-calling separated from tool implementations in relay/tools.js

deep_research_agent/research_manager.py: separate agent-style components (planner, search, writer, email) showing orchestration patterns

Security & Auth

5/10

Protecting data and access

Security awareness is visible: device OAuth flows, .env guidance, budget/billing latch, file permission hygiene; input validation and hardened auth lifecycle are pragmatic but not exhaustive in the public code.

Evidence

grokbot_tv_launcher/relay/README.md: explicit guidance to copy .env.example and chmod 600 .env (secrets hygiene)

grokbot_tv_launcher/relay/brain-llm.js: checkBilling() and BUDGET_FILE persisted with saveBudget() including 0o600 file mode usage

grokbot_tv_launcher/js/discover.js: Trakt device flow with refreshIfNeeded(), verify(), and localStorage guarded handling of tokens

Reliability & Observability

6/10

Stability and monitoring

Strong reliability and observability practices for the project size: metrics, budgeted LLM calls, transient error handling with retry/wait heuristics, cancellable contexts, and comprehensive end-to-end tests that exercise failure modes.

Evidence

grokbot_tv_launcher/relay/metrics.js: begin/timeTool/record/aggregate used to surface timing and usage

grokbot_tv_launcher/relay/brain-llm.js: transient error detection, retryDelay parsing, cancelable ctx, and budget counting with daily caps

grokbot_tv_launcher/tools/test-phone.js and grokbot_tv_launcher/tools/test-nav.js: extensive headless integration tests covering timeout, fallback, pairing, and offline flows

Expertise

Backend AI & LLM· Senior

Node.js· Middle

Python· Middle

Messaging & Real-time· Senior

Microservices & API Architecture· Senior

Industries

Media & Entertainment· Senior

Internet Services· Senior

Technologies

Python· Senior

Node JS· Senior

Flask· since 2015

Asyncio

Pydantic

Requests

Dotenv

Recommendations

Develop LLM-backed consumer features that require tool-calling and safe fallbacks (voice assistants, play-on-TV flows, recommender connectors).
Build resilient edge relays and device integrations (SSE/WS, pairing flows, Wake-on-LAN, robust STT fallbacks) for home/IoT products.
Implement production-grade LLM budgeting, observability and incident runbooks for teams adopting paid LLMs (billing latches, token accounting, fallback policies).
Lead development of integration and end-to-end test suites that exercise failure modes (network, transient LLM errors, device pairing), as already started here.

Repositories

The developer's experience in this domain has been verified based on AI analysis of the following repositories:

grokbot_tv_launcher ai_portfolio deep_research_agent

Senior Frontend Developer Confidence: High UI Engineer

UI Component Architecture

5/10

How interface parts are built

Component boundaries are pragmatic and hand-rolled for a device-first UI; many reusable UI utility modules exist but there is no modern component framework or bespoke design-system library extracted into a package.

Evidence

grokbot_tv_launcher/js/app.js: central UI composition and page lifecycle (boot, paintTabs, buildDock, renderPicks)

grokbot_tv_launcher/js/voice.js: encapsulated voice UI lifecycle (build, renderSkeleton, renderAnswer)

grokbot_tv_launcher/js/tvcard.js: card/panel component logic (buildCard, paintCard, showPanel, hidePanel)

Responsive & Cross-browser

6/10

Works on all screens and browsers

Strong cross-device and cross-browser attention with explicit mobile PWA meta, responsive CSS, and feature-detection fallbacks for Safari/iOS; not an advanced CSS-system but pragmatic and robust.

Evidence

grokbot_tv_launcher/relay/phone.html: PWA meta tags, viewport/mobile web app capabilities and touch-action handling

dynamic_pricing_engine/reports/pricing-research/index.html: responsive layout using CSS grid and clamp(), ARIA-friendly tablist and print styles

Performance Optimization

5/10

Speed of the interface

Measured and practical performance awareness (frame measurement, FPS toggle, throttling/repeat control) and attention to not piling network requests; lacks full bundle/CI performance artifacts but shows targeted optimizations.

Evidence

grokbot_tv_launcher/js/app.js: measureFps and toggleFps utilities

grokbot_tv_launcher/relay/phone.html: holdable() with controlled auto-repeat and press-throttling to avoid piling up remote/key requests

grokbot_tv_launcher/js/play.js (usage pattern): careful remote/network request gating (ensureOn / tryLaunch patterns)

Accessibility & Semantics

6/10

Usable for everyone

Accessibility and semantics are intentionally present: keyboard handling for custom widgets, ARIA roles/controls, and live regions; these are real, not purely copied, and include focus/keyboard shortcuts for tab lists.

Evidence

dynamic_pricing_engine/reports/pricing-research/index.html: tablist with role=tablist, keyboard handlers (ArrowLeft/ArrowRight/Home/End) and aria-controls

grokbot_tv_launcher/relay/phone.html: role and aria-label usage across interactive elements (dialogs, canvas alt text, live status), accessible hold-to-talk hints

State Management & Data Flow

7/10

Managing data in the app

Robust async state discipline and data-flow handling for complex, multi-source voice flows: request cancellation, restart-on-failure, fallback strategies, and deterministic test coverage for race conditions.

Evidence

grokbot_tv_launcher/relay/phone.html: voice/STT state machine (holdStart, startRec, recEnd, finish) with retries, abort handling, 'stale' result handling and mode switching between local Web Speech and server STT

grokbot_tv_launcher/tools/test-phone.js: Playwright-based tests that simulate iOS/Safari recognition quirks and validate restart/retry/cancellation behaviors

UX & Visual Polish

7/10

Look and feel quality

High UX polish for device-specific experiences: custom touch/mouse affordances, nice glass-style visuals, skeletons and meaningful states (loading/error/empty), and product-oriented interactions like Show on TV with two-step confirmation.

Evidence

grokbot_tv_launcher/relay/phone.html: polished CSS, mic UI states, listen/send/cancel UX and 'Show on TV' two-tap behavior

grokbot_tv_launcher/js/voice.js: renderSkeleton and structured assistant bubble rendering including stats and played-chip UI

dynamic_pricing_engine/demand_side_dashboard.html: charted dashboards with annotated tooltips and accessible canvas ARIA labels

Expertise

React· Junior

Modern Web Frameworks· Middle

Frontend Architecture & Build Tools· Middle

Industries

Media & Entertainment· Middle

Transportation & Logistics· Middle

Technologies

JavaScript· Senior

Tailwind CSS

Next.js

Chart.js

React.js

Vite

ESLint

Systems· mentioned only

UI· mentioned only

Recommendations

Lead development of interactive TV and companion-phone PWAs that require robust voice, TTS and STT flows and cross-device fallbacks.
Build and maintain device-specific UI surfaces where accessibility, keyboard navigation and non-pointer input are critical.
Implement and harden complex async UI state machines (voice, streaming, remote control) and their end-to-end test suites.
Prototype data-rich dashboards and visualization UI that integrate with backend experiment outputs and allow audited explanations.

Repositories

The developer's experience in this domain has been verified based on AI analysis of the following repositories:

dynamic_pricing_engine grokbot_tv_launcher ai_portfolio

Job preferences

Work permits: United States
Employment type: Full-Time
Locations: San Jose

Open to AI agents: they read this profile and start a dialog through Alionalion.io/@rudraksh-syal-2/mcp